►
From YouTube: Lightning Talks: William Buchwalter: Microsoft Azure
Description
Lightning Talks: William Buchwalter: Microsoft Azure
A
All
right
there,
you
go
hello
everyone,
so
my
name
is
William
Bush
Walter
I'm,
the
senior
software
engineer
at
Microsoft
in
the
a
I
own
research
group.
So
just
to
give
you
a
bit
of
context,
I'm
not
going
to
talk
about
Asher,
mostly
just
commerce
in
general
and
I've,
been
working
in
the
kubernetes
/ml
space
for
the
past
18
months,
I've
actually
been
committed
to
cool
off
since
last
July.
It
wasn't
called
cool
back
then,
but
still
and
a
bunch
of
other
stuff.
So
I
just
want
to
talk
a
little
bit
about.
A
Why
are
we
interesting
in
kubernetes
interested
in
kubernetes
for
machining?
In
the
first
place,
right
kubernetes
has
been
developed
with
micro
services
in
mind,
not
GPU,
workloads
or
anything
like
that.
So
why
does
it
make
sense
to
use
kubernetes?
Obviously,
the
the
biggest
the
strongest
point
for
community
is
the
community
right.
This
community
is
just
amazing
and,
and
so
large,
that's
if
you're
a
company
wanting
to
do
machine
learning,
training,
for
example-
and
you
want
apply
a
new
training
strategy,
something
let's
say
like
publishing
baserunning,
it's
actually
kind
of
complicated
to
do.
A
But
you
have
a
good
chance
of
finding
an
open
implementation
already
working
for
you
on
kubernetes.
So
obviously
this
is
stronger
argument,
but
then
it's
also
because
kubernetes
I
think
is
really
well
designed
and
clean
api's.
So
that
means
even
if
you
don't
find
what
you
want
and
you
need
to
start
from
scratch.
A
It's
actually
much
easier
that
unco
Burnett
is
then
it
was
just
a
few
years
ago,
for
example,
I
worked
actually
on
population-based
training
so
which
come
from
deep
I'm,
originally
with
a
large
customer
and
and
to
implement
that
on
kubernetes
it
just
took
a
few
days
and
another
shot.
It's
actually
really
easy,
because
the
API
is
already
nice
and
obviously
scaling
is
important.
Kubernetes
can
scale
pretty
loudly.
So,
for
example,
we
have
a
nice
class
to
deal
with
open
area.
A
So
a
few
months
ago,
I
think
in
January
opening
I
released
this
blog
post
called
scaling
kubernetes
to
2500
nodes,
so
detail
measure,
and
you
know
it
wasn't
easy.
They
had
a
lot
of
issues
with
EDD
network.
This
guy,
you
excetera,
but
ultimately
they
managed
to
use
a
scale
with
a
very
small
team
of
engineers.
I
think
they
were
two,
maybe
three
people
and
a
single
job.
A
In
that
case
10k
course
that's
pretty
big
you-
and
this
was
the
finishes
flutter
like
last
year
or
three
years
ago,
and
with
every
single
release
of
kubernetes
and
HCD,
it's
becoming
easier
and
easier
to
go
even
further
than
that.
Some
really
excited
to
see
where
this
is
going
yeah,
that's
my
other
slide,
I
guess
so
we
have
kind
of
two
offering
for
kubernetes
on
Azure.
We
have
a
key
s,
which
is
the
full
manage
kubernetes.
A
So
I'm
not
going
to
talk
about
everything
here,
but
I
won't
talk
about
two
things
that
I
think
are
going
to
be
interesting.
So
it's
a
bit
far-fetched,
but
the
first
one
is
visual
couplets.
So
if
you
didn't
know
about
that,
that's
a
project
basically
to
an
open
source
implementation
of
the
couplet
that
you
can
then
back
up
with
usually
something
like
Azure
contra
instance
or
AWS.
But,
for
example,
someone
just
made
a
request
to
other
provider
for
al-jabat.
A
So
as
your
batch
lets,
you
run
basically
GPU
jobs
and
you
might
wonder
why
you
want
to
do
that
instead
of
just
using
GPU
in
communities.
The
reason
is
because
you
can
scale
very
fast
in
in
a
matter
of
seconds,
with
a
low
batch
and
so,
for
example,
it
would
be
very
nice
for
use
cases
when
you
want
to
do
it
transfer
running
on
very
short
running
times
and
when
you
want
to
keep
control
of
the
cost
another
one
which
I'm
excited
about,
but
it's
very
early
is
meta
particular.
A
So
if
you
were
at
cook
or
nest
you're
in
Austin,
you
might
have
seen
the
keynote
by
Brandon
downs.
Where
basically
made
this
point,
that
kubernetes
is
becoming
the
standard
runtime
of
the
cloud
right
and
since
it's
the
runtime,
we
also
need
a
standard
library
to
go
with
it.
So
you
can
directly
from
your
code,
deploy
to
kubernetes
without
having
to
go
for
Duke
and
kubernetes
templates,
and
so
I
mean
I'm
playing
with
this
idea
of
tailoring
meter
particular
to
work
specifically
for
machine
learning.
A
So,
for
example,
you
could
define
a
decorators
in
Python
on
top
of
your
function,
to
say:
okay,
I
want
to
train
this
function
using
that
mini
agent
in
parallel
exit
era,
and
when
you
do
Python
my
script,
it's
actually
going
to
deploy
everything,
build
everything
and
deploy
on
the
cloud
for
you,
for
example,
using
control-c
Rd,
some
Charlotte,
so
obviously
extremely
experimental,
but
you're
just
sharing
a
few
files
that
I
think
are
interesting
and
that's
it
for
me.
Thank
you.