►
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
Birds of a Feather: Do You Really Need on Premises Serverless ? - Igor Khapov, IBM
Speaker will describe serverless technology, most popular use cases and how international companies use serverless in the public and private clouds.
Serverless can be used on premises and in the cloud. It can help you to create the most cost effective model. Serverless on premises provides you ability to utilize your own resources.
Speaker will also demonstrate how to create infrastructure with multiple architecture support for serverless application and describe situations when you need it.
https://sched.co/MPeX
A
Hello,
everybody
and
actually
I-
already
saw
that
a
couple
of
guys
ask
me
where
part
of
my
presentation
name
Joseph,
is
er
and
actually
built
a
theater.
It's
type
of
a
presentation
of
the
CN
CF.
This
mean
that
it
should
be
like
more
discussion,
and
if
you
have
any
questions,
we
can
discuss
it
later
at
the
ends
of
the
presentation
or
you
can
try
to
interrupt
me.
A
Okay,
my
name
is
the
good
cop
of
I
am
developer
in
the
Moscow
Science
and
Technology
Center.
It
is
development
lab
and
traditionally
we
are
creating
IBM
software.
We
participated
in
the
different
projects
related
to
the
cloud
to
on-premises
solution,
and
so
on.
One
of
my
last
project
is
Watson
studio.
Watson
studio
is
cloud
and
on-premise
solution
to
simplify
data
sent
each
day
we
have
the
Peters
opinion
our
studio.
We
work
with
GPU
devices
and
the
whole
production
work
on
top
of
the
kubernetes,
and
when
we
created
this
infrastructure
we
connected
some
issues.
For
example.
A
This
means
that
we
are
talking
about
design
where
we
have
some
back-end
code
hosted
in
the
cloud
hosted
on
premises
and
we
have
some
functions
traditionally.
This
function
has
input.
Output
contains
stateliest
working,
the
stateless,
and
traditionally
it
is
Hardware
agnostic,
IBM
joined
to
this
history
about
four
years
ago
and
actually
right
now
there
are
a
lot
of
serious
frameworks
exist.
A
There
are
a
lot
of
cloud
serious
classrooms.
There
are
a
lot
of
solutions
which
can
be
installed
on
premises,
IBM
contribute
to
the
open
risk
projects,
it's
a
it's
an
apache
license
and
it's
written
on
the
scala
and
today
damn
I
will
show
you
how
I
use
it
for
my
use
cases
each
time
when
I
am
working
with
different
development
teams,
when
I
am
talking
with
the
clients,
everybody
asked
me:
what
is
the
best
use
cases
to
use
Ciara's?
A
A
Okay,
and
my
story
in
the
Sierras
was
maybe
began
to
two
years
ago
and
I
participated
in
the
hackathon
in
the
Bucharest
and
the
company
which
organized
this
hackathon
was
a
newspaper
company
and
the
goal
was
to
find
out
fake
news.
We
have
a
feed
of
the
news
and
we
should
determine
if
it's
fake
or
not,
and
during
the
hackathon
we
created
no
J's
application
weekly
and
who
know
what
close
out
meals
go
down.
It's
coach,
DB
but
hosts
in
the
closets.
A
It
was
separated
company,
but
IBM
bought
it
maybe
five
years
ago,
if
you
will
see
close
down
this
means
is
it
can
be
replaced
by
a
coach
to
be
ok
and
we
decided
to
use
serious
technologies.
You
know
from
my
team
implemented
the
solution
where
we
get
the
Indians.
If
we
have
a
new
news
in
the
feed
and
we
created
a
function
which
just
connect,
the
topic
connect
some
statistic,
information
and
after
that
we
put
everything
to
the
Wharton
studio.
What
estudia
contains
model
model
management
model,
training,
information
for
hackathon?
A
We
just
created
endpoint
for
our
models
and
we
parse
the
whole
article
to
machine
learning
and
if
we
are
talking
about
serials
and
if
we're
talking
about
open
risk,
it's
even
driven
what
that
means.
We
have
a
packages
which
can
provide
different
events
based
on
these
events.
It
can
be
collected
by
the
triggers
and
we
have
set
of
rules
which
determined
which
function
should
be
started
and
if
function
should
be
started.
After
that
we
have
an
action.
A
In
our
case,
it
will
be
docker
container,
hosted
on
premises
and
the
result
of
the
function
is
JSON
which
will
be
stored
in
the
coach
dB.
Let's
look
at
the
architecture
of
the
its
common
architecture
of
the
open
risk.
We
haven't
jinks.
We
have
controller
controller
responsible
to
Minotaur
all
and
workers
for
scheduling
the
function
between
different
clusters
or
in
workers.
I'm
talking
about
clusters
and
workers
today,
I
will
show
them
where
we
will
use
a
single
kubernetes.
Cluster
single
kubernetes
cluster
contains
only
one
in
worker.
A
This
means
that
this
in
worker,
tied
with
kubernetes
editor
and
can
invoke
a
lot
of
function
inside
the
kubernetes.
But
if
we
want
we
can
we
can
attach
methods,
we
can
attach
no
occur
and
as
a
cluster
system
to
this
architecture,
and
all
results
will
be
stored
to
the
coach.
Db
I
know
about
some
experiments
when
people
was
not
satisfied
to
store
or
result
results
in
JSON
format
to
the
coach
DB.
For
that
purpose,
we
can
replace
the
coach
DB
v
is
in
the
memory
database.
A
It
is
not
easy
but
possible
and
of
course,
we
have
a
Kafka
which
will
distribute
the
actions
to
the
different
ten
workers.
The
only
one
point
that
we
also
should
have
a
docker
registry,
which
will
provide
docker
containers
to
all
this
infrastructure,
ok
eyes
and
actually
open
risk
get
projects
contains
a
lot
of
information
collected
a
lot
of
projects,
the
first
portion
regarding
platform.
We
have
open
risk
itself
written
in
Scala.
We
have
common
line
interface,
we
have
a
PA
gateway
and
we
have
catalog
where
we
can
install
our
packages.
A
We
also
have
a
lot
of
runtimes.
This
means
that
if
you
want
to
write
down
a
function
based
on
node.js,
Petone
and
other
languages,
you
can
just
get
this
runtime
and
use
it,
but
why
I
decided
to
use
open
risk
for
my
use
case,
because
there
is
a
possibility
to
create
your
own
liquor
container
and
integrate
all
right
libraries
in
sizes
that
docker
container?
A
We
also
have
a
lot
of
tooling.
For
example,
we
can
deploy
open
risk
using
ansible.
We
have
a
help
chart
to
deploy
the
whole
system
open
risk
on
top
of
the
kubernetes.
It
can
be
installed
on
the
vehicle
motion
of
the
hard
van,
so
on
and
so
on,
and
of
course,
you
can
add
in
workers
wherever
you
want,
we
have
a
set
of
packages.
Packages
means
that
you
can
create
your
own
package
and
consume
and
produce
events
from
your
own
application.
It
can
be
just
cool
post
the
process.
A
A
Okay,
after
that
I
decided
to
investigate
what
what
is
these
cases
applicable
to
the
data
science
from
Searles
air,
the
first
one,
what
we
did
during
the
hackathon?
It's
just
reaper
testing
the
data,
it's
really
suitable
and
the
common
feature
of
the
Sierras
when
we
just
process
the
data
and
after
that
start
may
be
data
science
store
to
the
database
and
so
on,
and
so
on.
I
find
out
the
team
from
IBM
Zurich,
which
tried
to
create
something
from
my
opinion.
A
Of
course,
functions
should
be
multi-tenancy
and
it
is
a
really
problem
to
store
the
state
if
you
have
absolutely
stateless
technologies.
For
that
purpose,
they
decided
to
use
a
batch
creo.
They
created
executors,
which
work
with
different
languages
and
they
created
a
driver
which
try
to
execute
functions
closer
to
the
data
close
to
the
edge
Crale,
and
they
are
trying
to
move
the
data
in
the
page
cradle
closer
to
the
function.
In
my
case,
I
decided
to
create
something
like
a
mix.
A
A
I
have
to
use
standard
flow,
the
enter
flow
state
and
it's
possible
to
stores
the
state
in
the
tensor
flow,
put
it
on
the
cloud
on
tour,
cause,
DBE
or
maybe
object,
storage,
and
after
that,
restore
it.
It
is
slow,
but
for
some
base
cases
when
you
should
support
multiple
versions
of
the
models.
It's
reasonable
on
this
slide.
I
draw
the
target
architecture
which
I
want
to
create:
it's
not
absolutely
radium.
A
We
have
API
gateway.
We
have
the
functions
which
is
responsible
to
collect
the
data
from
different
sources.
We
have
a
function
which
is
responsible
to
train
the
model,
and
we
have
something
like
check
points
which
store
the
state
of
the
model
to
the
different
in-memory
database,
to
storage
and
so
on.
It
depends
on
say
his
case
and
depend
on
which
client
want
to
have
such
infrastructure
and
what
I
really
like
that
using
functions,
we
can
implement
the
Train
model
based
on
the
open
risk
event.
If
something
was
changed.
A
For
example,
if
we
have
a
new
feed
for
training
for
retraining
the
data
which
we
just
get,
this
event
returns
the
model
stores
the
state
and
after
that,
deploy
the
production
model.
Ok
and
actually
I
work
with
GPU
devices.
On
top
of
the
power
architecture
who
know
what
power
hitter
two
means
a
couple
of
guys
and
actually
it's
something
like
Raspberry
Pi,
but
open-source
technology
for
architecture,
which
contains
a
lot
of
benefits
for
some
use
cases.
A
It's
this
is
speech
not
about
power,
but
you
should
understand
that
it
is
absolutely
different
architecture
and
you
should
build
source
code
and
docker
images
separately
and
actually
I
decided
to
use
power
for
model
training,
because
it's
quite
powerful
I
decided
to
use
extra
texture
for
common
goals
and
I
decided
to
use
JSON
devices
to
run
the
model
who
know
what
Jets
in
the
race,
yeah
and
actually
just
on
device.
It's
a
media
platform
or
something
like
a
Raspberry
Pi,
but
with
integrated
GPU
device.
A
I
bought
dredged
jets
on
Nana
and
try
to
integrate
it
in
the
cluster
right
now.
This
means
that
it
is
quite
cheap
device,
just
another
prize
about
100
dollars,
but
it
with
GPU,
and
if
you
want
to
analyze
the
video
stream,
you
shouldn't
stream
the
whole
video
to
the
Internet
to
the
cloud
you
can
analyze
this
video
in
place.
A
Okay
and
if
I
have
a
mesh
of
architecture
and
if
I
want
to
create
a
docker
images
for
which
will
be
started
on
the
all
of
this
architecture.
I
should
create
dr.mini
first,
who
heard
about
docker
manifest
okay
and
actually
dr.
manifest,
provide
us
possibility
to
create
to
build
docker
images
for
exactly
texture
for
power
architecture,
for
ARM
architecture
and
after
that
creates
something
like
a
will
show,
which
should
local
image,
which
will
aggregate
all
these
architectures
containers
images
and
when
docker
will
try
to
pull
these
docker
images.
A
Okay,
of
course,
it
is
not
working
from
the
box
from
the
open
risk
because,
because
of
the
it
doesn't
contains
possibility
to
specify
the
node
where
you
want
to
start
your
function
and
as
I
already
mentioned,
it's
written
in
Scala
and
I
was
a
Java
developer
five
years
ago
and
when
I
changed
the
course
I
realized
why
I
decided
to
use
node.js
fight
and
so
on.
There
are
a
lot
of
level
of
abstraction
and
open
risk
work
with
methods
with
no
occur
with
kubernetes.
A
The
master
on
top
of
the
X
architecture
and
I
have
two
power
workers,
but
enable
scheduling
for
the
master,
which
will
mean
that
we
will
be
able
to
start
function
on
the
two
architecture,
power
and
acceleration.
Okay
and
I
want
to
show
you
also
how
to
I
did
that.
Okay
I
already
mentioned
that
we
are
able
to
create
customized
actions
based
on
the
docker
image.
A
Of
course,
if
I
am
using,
multiple
architectures
I
should
build
it
from
scratch
in
the
github
I
find
out
the
local
file
which,
with
the
name
docker
skeleton,
the
car
skeleton,
provide
a
speciality
to
work
with
open
risk
and
yeah,
and
actually
you
can
use
any
base
image,
but
it
should
contains
the
flask
and
it
should
contains
action
proxy
action
proxy.
Just
a
simple
implementation
of
what
function
should
do
and
if
we
will
go
to
the
source
code
and
close
what
we
do
not
need.
It
contains
information
about
how
we
should
emit
this
function.
A
This
means
that
when
we
create
the
function,
init
function
will
be
started
inside
the
docker
container.
For
example,
if
you
want
to
provide
your
container
to
be
ready
for
input
and
attach
some
additional
resources,
it
should
be
done
on
the
unit
state
and
we
have
run
and
this
function
will
be
started
when
we're
invoking
the
function.
A
Have
a
docker
image
for
power,
he
texture
for
exact
idea
and
the
lattice.
It's
not
traditional
docker
image.
It's
manifest,
which
point
me
to
both
of
the
images
I
pushed
it
on
my
own
registry
and
in
this
cluster
I
also
have
installed
open
risk.
For
example,
you
can
have
separated
installation
of
the
open
risk
and
you
can
install
separated
clusters
for
in
workers,
but
I
mix
infrastructure
to
reduce
some
resources,
I
install
it
in
the
separated,
namespace
open
risk.
A
Okay,
as
you
remember
picture,
we
have
API
gateway.
Here
we
have
a
lot
provider
which
listening
different
events.
We
have
coach,
DB
and
code
and
provider
which
contains
the
created
functions
and
result
of
that.
We
contains
a
controller
which
decide
where
should
be.
Where
should
we
start
the
function?
And
if
we
have
three
function
or
should
we
invoke
some
additional
ports?
If
we
are
overloaded,
we,
of
course
we
have
a
Kafka
or
we
haven't
Jenks.
A
A
A
Okay
and
if
we
will
go
right
now
to
the
open
risk
name,
space
right
now,
not
simple
happen,
because
I
do
not
specify
that
I
require
provide
container.
This
means
that
container
will
be
created
only
after
the
first
invocation
of
the
function,
and
after
that
the
action
was
already
created
and
invoked.
I
want
to
invoke
it.
A
B
A
A
For
the
result
of
this
function,
I'm
storing
results
to
the
separated
environment,
I
am
just
showing
which
architecture
I
used
to
start.
This
function,
you,
if
you
have
a
small
piece
of
data
which
your
function,
values
or
you
should.
If
you
want
to
create
a
chain
of
functions,
you
can,
for
example,
create
a
lot
of
JSON
attributes
and
provided
from
one
function
to
another
function,
to
store
to
the
database
and
so
on
and
after
that,
I
want
to
delegate
this
function.
I
want
to
create
the
function
on
the
different
architecture.
A
A
And
it
is
based
on
the
exergy
tech
show
it
will
provide
us
possibility
not
only
to
share
to
use
different
architecture
in
the
same
cluster.
For
example,
if
you
are
using
the
GPU
devices
you
can
attach
in
worker
directly
to
the
docker,
and
in
that
case
all
actions
which
will
be
started
using
this
worker
will
be
able
to
share
GPU
devices
and
it
will
provide
possibility
to
increase
the
utilization
of
the
GPU
devices
of
the
bagua
and
so
on,
and
so
on.
A
Okay
and
why
I'm
doing
that,
because
I
have
not
absolutely
satisfied
how
we
are
utilizing
our
resources,
for
example,
if
I
want
to
create
a
pipeline
for
my
data,
scientist
I
have
I
should
have
dedicated
resources
for
this
team,
and
this
team
could
not
be
shot
at
with
different
departments,
because
it
will
be
really
difficult
to
share
GPU
devices
to
collect
the
whole
information.
But
if
we
will
try
to
apply
the
CAS
technologies
for
this
use
cases,
we
can
get
more
results.
A
For
example,
it's
a
real
problem
to
create
a
function
calculated
to
your
data
storage
in
the
IBM.
We
have
a
project
where
all
data
storage
ends
up
eclipse',
but
I
have
no
enough
resources
and
PI
keeps
you
to
start
all
data
science
project
analyze.
These
results
and
I
have
to
start
it.
Maybe
in
London
the
Amsterdam
and
it's
quite
big
delay
to
download
this
data
to
the
another
one
region
to
analyze
it
upload
results
and
so
on
and
so
on.
For
some
functions.
It
is
really
important
to
have
more
resources
there.
A
Additionally,
for
each
functions
we
have
timeout.
We
have
limitation,
for
example,
that
function
should
be
completed
in
one
minute,
that
it
should
be
only
with
one
gigabyte
of
RAM
and
so
on
and
so
on,
and,
to
summarize
everything
I
suppose
to
understand.
If
you
really
need
on-premise
service,
you
should
have
a
lot
of
in
company
developers,
because
it's
require
a
lot
of
development.
A
A
I
worked
with
the
data
scientist
and
they
liked
this
project
because
they
I
they
have
to
split
their
huge
pipeline
into
small
pieces,
and
it
is
quite
easy
to
support
small
pieces
of
the
data
science
cord,
then
to
have
a
mesh
of
the
everything
and,
of
course
you
should
have
a
time
to
implement
everything,
because
right
now,
I
have
no
any
idea
which
framework
supported
from
from
the
box
but
I
hope
to
upstream
the
water.
What
I
did
to
the
open
risk
and
hope
to
it
will
be
available
for
everybody.
Thank
you
very
much.
A
A
It
will
be
something
like
an
edge
when
we
will
have
a
lot
of
devices
directly
to
the
video
cameras
to
analyze
the
video
stream,
and
we
will
have
cloud
solution
which
will
train
the
model
and
the
second
solution,
which
already
badly
implemented
and
will
be
in
the
production
I
suppose,
is
to
use
cierra
less
not
to
the
whole
dataset
cycle,
but
just
around
video
analytics,
it's
already
builded
and
prepared,
prepared,
doaker
containers
and
which
does
should
start
it.
Onyx
architecture.