►
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
Kubernetes the New Research Platform - Lindsey Tulloch, Brock University & Bob Killen, University of Michigan
Academic research institutions are at a precipice. They have historically been constrained to supporting classic “job” style workloads. With the growth of new workflow practices such as streaming data, science gateways, and more “dynamic” research using lambda-like functions, they must now support a variety of workloads. In this talk, Lindsey and Bob will discuss some difficulties faced by academic institutions and how Kubernetes offers an extensible solution to support the future of research. They will present a selection of projects currently benefiting from Kubernetes enabled tools, like Argo, Kubeflow, and kube-batch. These workflows will be demonstrated using specific examples from two large research institutions: Compute Canada, Canada’s national computation research consortium and the University of Michigan, one of the largest public Universities in the United States.
https://sched.co/MPY6
A
All
right,
good
afternoon
we're
going
to
start.
The
third
talk
in
the
case
studies
track
kubernetes
the
new
research
platform
given
by
Lyndsey
and
Bob
pretty
interesting,
especially
when
you
think
about
what
research
institutions
have
to
go
through,
especially
about
offering
up
compute,
so
that
researchers
can
actually
do
the
work
that
they
need
so
be
sure
to
go
into
the
app
and
rate
this
session
afterwards.
Also
we'll
take
questions
at
the
end
as
well,
and
I'll
run
around
with
a
microphone
at
the
end,
so
that
everyone
can
hear
and
also
it
is
being
recorded.
B
C
Hi,
my
name
is
Bob
Cohen
I
am
a
senior
research
cloud
administrator
for
the
University
of
Michigan.
I
am
also
a
cloud
native
computing
Foundation
ambassador
with
a
focus
in
academic
outreach,
and
this
is
kubernetes
the
research
platform
or,
as
I
really
like
to
think
about
it,
a
tale
of
two
research
institutions.
So
there
will
be
a
little
bit
of
back
and
forth
in
this
presentation
in
this
case
study,
because
it's
sort
of
how
we
gained
traction
or
adoption
of
kubernetes
at
both
our
institutions.
C
We
both
sort
of
came
to
the
same
conclusions,
encountered
some
of
the
same
problems
that
sort
of
dig
into
like
why
so
at
both
of
our
institutions,
we've
seen
an
increased
usage
of
containers
well
everywhere.
I
think
that's
pretty
safe
to
say
in
the
general
research
community
as
well,
and
this
isn't
just
like
what
we
normally
think
of
as
containers,
but
also
through
other
sort
of
like
runtimes,
like
singularity
charlie
cloud.
And
although
those
are
you
know,
more
meant
for,
like
HPC
systems,
they're
still
containers
and
still
sort
of
relevant
to
us.
C
But
one
thing
we've
really
seen
is
a
general
movement
away
from
sort
of
the
strict
style,
job
workflow
and
a
greater
need
to
sort
of
support
things
like
data
streaming
services
with
like
and
with
in-flight
processing
or
eventing.
So
a
lot
of
things
like
Kafka
spark,
flink
and
just
general
dependence
on
more
sort
of
like
persistent
services,
things
like
databases,
other
things
sort
of
like
interactive
science
gateways
like
Jupiter
hub,
and
these
are
things
that
you
know
don't
classically
mesh
well
with.
You
know,
HPC.
C
Know
the
surprises
why
kubernetes
well
both
are
groups
again
sort
of
came
to
the
same
conclusion
and
that
career
neighs
has
become
these
sort
of
standard
for
container
orchestration
like
it
or
not?
It
really
is,
and
the
big
thing
is
what
it
can't
do.
It
can
be
extended
to
do
and
if
it
can't
be
liked-
and
it
can
all
be
done,
while
essentially
avoiding
vendor
lock-in.
C
So
what
I
do
is
reproducible
by
someone
else
that
can
get
a
Cabrini's
institution
up,
kubernetes
installation
up
and
going,
and
so
it's
in
a
big
boon
for
reproducibility
in
science.
You
don't
have
to
change
anything.
If
you
have
a
conformant
kubernetes
cluster,
if
I
can
run
this
here,
it
should
be
able
to
run
there.
C
There's
been
a
lot
of,
like
other,
just
general
traction
in
the
cabeza
ecosystem.
You
know
there's
about
8,000
attendees.
Here
we
got
66,000
people
in
the
kerb,
Brunetti
slack
and
well
just
a
massive
amount
of
adoption,
but
with
that
little
bit
out
of
the
way,
I
actually
want
to
hand
it
back
over
to
Lindsey
to
sort
of
go
over
things
from
a
research
perspective
and
I'll
cover
things.
A
little
bit
later,
as
more
of
a
service
provider
perspective.
B
Perfect,
okay,
so,
as
I
mentioned,
I
am
just
finishing
up.
My
Bachelor
of
Science
in
computer
science,
at
Brock,
University
and
Brock
offers
a
opportunity
to
do
a
work
placement
as
part
of
your
degree,
so
I
actually
spent
16
months
as
an
intern
at
Red
Hat
working
on
multi
cluster
kubernetes,
and
that
was
kind
of
my
introduction
into
the
kubernetes
community
and
when
I
was
returning
back
to
school
in
September
I
had
a
few
more
credits.
B
To
finish
up
so
I
was
thinking
how
can
I
finish
up
my
degree
and
still
continue
learning
about
kubernetes
and
so
I
decided
to
approach
my
previous
research
supervisor
that
I
had
done
a
research
project
under
before
and
ask
for
her
advice
about
how
I
could
incorporate
kubernetes
into
a
research
project,
and
my
previous
research
was
under
dr.
Sheridan
Houghton.
B
Optimization
problems,
in
other
words,
is
an
algorithm
that
has
a
bunch
of
tunable
parameters
and
a
memory
of
best
path
that
is
passed
through
a
bunch
of
iterations,
with
the
aim
of
solving
computation,
computationally
hard
problems
in
a
really
hockey
but
expedient
way,
and
it
can
be
configured
to
give
significant
results
for
the
DNA
fragment
assembly
problem.
And
this.
B
But
there
was
a
small
catch
and
that
was
to
try
and
get
it
up
and
running
on
compute
Canada
at
the
time
I
had
not
heard
of
compute
Canada
I
suspect.
Many
of
you
also
don't
know
about
compute
Canada.
So
let's
learn
about
compute
Canada
together,
all
right,
so
compute
Canada
operates
at
a
national
level
through
partnerships
with
regional
organized
patience.
So
there's
a
snot
on
the
east
coast
of
Canada
it'll,
Atlantic
Canada,
calcul
Quebec,
which
operates
in
Quebec
compute,
calcul
Ontario
in
Ontario
and
West
grid,
which
covers
the
west
coast.
B
Ok,
so
compute
Canada
is
appealing
to
many
Canadian
researchers,
because
all
Canadian
faculty
members
have
access
to
compute
Canada
systems
and
are
able
to
sponsor
their
students
and
postdocs
for
no
fee.
So
this
provides
greater
compute
power,
expertise
and
resources
that
are
typically
available,
especially
at
smaller
research
institutions
that
have
limited
funding
and
resources,
and
this
data
is
from
2016,
but
it
kind
of
gives
an
overview
of
the
different
types
of
research
that
compete.
Canada
is
being
used
for
as
far
as
resources
go
compete.
B
Canada
has
compute
and
storage
resources
spanning
across
five
to
ten
data
centers
with
300,000
cores,
twelve
peda
flops
and
50-plus
petabytes
of
storage.
There's
a
team
of
about
200
experts
with
at
least
one
place
locally
at
each
institution
across
Canada,
and
these
experts
help
out
with
the
installation
and
configuration
of
hundreds
of
research
software
packages.
B
Most
production
work
currently
is
done
as
batch
jobs
on
compute
nodes
via
the
slurm
scheduler
and,
as
bob
alluded
to
earlier
compute.
Canada's
user
base
is
broad
and
it's
bringing
in
a
broader
set
of
needs.
So
there's
currently
a
tremendous
interest
in
services
enabling
research
data
management
and
other
things
that
kubernetes
is
really
good
at
solving
alright,
so
back
to
the
salmon
algorithm.
B
So
now
we
know
a
bit
of
a
compute
Canada,
but
it's
still
not
totally
clear
how
I'm
going
to
get
the
salmon
algorithm
running
on
come
on
kubernetes
on
their
resources.
As
far
as
I
can
tell,
through
my
contact
with
the
compute
canada
expert
at
Brock,
there's
no
kubernetes
cluster
that
is
currently
up
and
running
on
compute,
canada,
resources
and
no
restrictions
on
researchers
doesn't
mean
they're
just
going
to
hand
over
admin
privileges
so
that
I
can
deploy
it
myself.
B
B
B
B
B
Oh
there
we
go
okay,
so
Alice
produces
several
petabytes
of
data
per
year
and
they
require
a
lot
of
compute
power
to
process
this
data
and
also
filter
it
before
it's
actually
made
into
tangible
results.
So
there's
a
Tier
one
compute
center,
that's
based
in
Vancouver
and
then
the
tier
two
competing
center
is
at
the
University
of
Victoria,
and
this
is
where
final
analyses
on
the
data
from
Atlas
are
performed.
So
the
UVic
Atlas
group
consists
of
25
scientists,
students,
professors,
technicians
and
computer
experts,
so
where
exactly
does
kubernetes
fit
into
all
of
this?
B
So
we're
currently
at
the
University
of
Victoria,
builds
off
of
previous
work
where
a
method
was
developed
to
deploy
a
kubernetes
cluster
as
an
Atlas
computing
site
and
some
of
the
details
you
can
see
from
the
image
the
panda
work.
Workload
management
system
used
by
Atlas
is
able
to
consume
kubernetes.
If
a
kubernetes
endpoint
is
provided
for
it,
and
then
it's
able
to
spin
up
Atlas
jobs
by
provisioning.
The
resources
on
the
computing
sites
they
have
access
to.
Cern
is
kind
of
the
poster
child
for
high
throughput
computing.
B
B
So
once
I
was
able
to
get
in
touch
with
the
compute
Canada
researchers
involved
on
the
west
coast,
things
got
a
lot
easier
because
they
were
a
lot
more
interested
in
what
I
was
trying
to
do
so.
I
was
able
to
get
on
our
Buddhists
cloud
project
access
to
an
our
Buddhist
cloud
project,
which
was
an
open
stock
instance
with
the
maximum
resource
allocation,
and
this
is
some
details
about
what
all
that
included.
B
B
B
B
So
since
I'm,
just
a
researcher,
I
can't
really
say
what
the
future
of
kubernetes
on
compute
canada
resources
will
be.
But
if
it's
in
line
with
the
stated
mission
mission,
sorry
researchers
like
CERN
are
driving
innovation,
and
this
is
true
with
kubernetes
as
well.
But
the
learning
curve
is
very
steep
and
researcher
time
is
precious.
B
So,
as
I
said,
I
spent
a
lot
of
the
time
that
I
was
working
on
this
project,
trying
to
just
figure
out
how
to
get
kubernetes
deployed,
rather
than
exploring
all
of
the
tools
that
I
could
use
if
I
had
a
kubernetes
cluster.
So
if
it
was
just
a
matter
of
learning,
gamal
and
bash
and
all
those
things
it
might
be
reasonable
for
a
researcher
to
to
learn
but
having
to
deploy,
kubernetes
and
learn.
C
C
So
we
happen
to
be
one
of
the
largest
universities
in
the
United
States,
with
19
schools
and
colleges
catering
to
about
45,000
students
and
roughly
8,000
faculty,
if
100
graduate
programs
in
the
top
10
according
to
US
News
and
World,
Report
and
sort
of
the
more
relevant
statistic
of
this
conversation
is,
we
haven't
be
number
one
in
research
volume
among
all
US
public
universities.
According
to
the
United
States
National
Science
Foundation
sort
of
backing.
That
up
is
we
have
just
under
about
1.5
billion
in
annual
research
expenditures.
C
So
we
are
essentially
tasked
with
handling
the
computational
research
needs
for
the
entire
university.
We
want
the
researchers
to
come
to
us
and
consume
these
sort
of
services
that
we
provide
instead
of
buying
a
computer
and
putting
under
their
desk
so
essentially
we're
a
our
pass
or
sort
of
research
platform.
As-A-Service
I
hate
that,
but
it's
kind
of
true,
so
we're
a
little
bit
of
infrastructure-as-a-service
a
little
bit
of
software
as-a-service
a
little
bit
of
container
as
a
service
now.
C
So,
if
backing
all
that
stuff,
our
group
is
roughly
25
people,
and
so
our
bread
and
butter
is
still
our
classic
HPC
and
we
have
27,000
cores.
We
also
have
a
restricted
data
side,
restricted
data
cluster
that
has
about
14,000
cores.
We
also
have
various
other
platforms,
both
arm
and
power,
whack
loads
of
GPUs.
We
have
various
data
science
initiatives
so
a
fairly
large
like
dhoop,
cluster
and
SPARC.
We
also
have
on
Prem
virtualization
services
and
because
research
grants
are
now
coming
with
cloud
credits.
C
So
supporting
all
these
different
systems
is
its
own
set
of
challenges
and
everything
that
was
on
Prem.
That
I
mentioned
generally
can't
get
out
to
the
Internet
for
security
reasons.
Ie,
we
don't
want,
like
a
compromised
account,
consuming
researchers
allocation
like
mine,
cryptocurrency,
or
something
like
that
not
to
mention
this
other
various
research.
That's
under
like
export
controls,
so
many
of
the
services
that
we
would
normally
consume
have
to
be
mirrored
internally
and
instead
of
standing
up
a
bunch
of
you
know,
virtual
machines
or
installing
on
bare
metal.
C
C
So
we
started
by
hosting
our
sort
of
ancillary
systems,
so
things
like
patch
mirrors,
build
systems,
source
repositories,
logging
and
monitoring,
along
with,
like
other
things
like
identity
management,
services
and
sort
of
once
we
cut
our
teeth
and
got
sort
of
all
the
kinks
worked
out.
We
decided
to
open
it
up
and
we
created
a
sort
of
container
as
a
service
offering
called
ODS
or
on-demand
services.
That
was
more
meant
for
our
researchers
and
when
we
gave
it
to
them,
we
learned
a
lot
of
things.
C
Well,
they
needed
a
sandbox
to
do
literally
anything
and
everything,
and
that
was
kind
of
crazy
and
when
they
need
to
do
with
like
being
able
to
stomp
on
each
other.
So
initially
we
had
many
requests
for
various
database
systems,
and
that
was
mostly
to
like
augment
jobs
that
were
running
on
our
HPC
cluster,
and
then
we
had
other
things
like
CV,
Methos,
c,
vm,
FS,
sorry
and
squid,
and
then
it
sort
of
grew
to
sort
of
custom
application
stacks
like
cloud
ER.
C
C
So
why
keep
flow
well
cube
flow,
bundles
up
a
slew
of
different
components
that
makes
them
easy
to
deploy
and
very
consumable
from
like
an
end
user.
Like
research
perspective.
Now
we
don't
actually
have
any
like
GPUs
in
our
current.
You
know
careers
offering
communion
cluster,
but
it
doesn't
mean
those
tools
aren't
useful.
A
good
example
here
is
the
MPI
operator,
so
that
can
be
used.
Other
things
like
ubers,
horev,
odd
model
for
like
training
and
ultimate
as
a
training
alternate
like
a
distributed.
C
Tensorflow
job
and
there's
also
a
whole
slew
of,
like
other
possibilities,
that
sort
of
fall
into
the
MPI
space,
not
to
mention
there's
other
things
that
are
more
broadly
usable.
That
can
be
like
run
anywhere
like
Argo.
The
big
thing
is
this
has
actually
gotten
researchers
and
also,
like
other
support
staff
and
more
people,
thinking
more
about
like
so
the
new
research
life
cycle
or
research,
software
engineering.
More
and
well,
this
graphic
is
sort
of
centered
around
machine
learning.
C
C
Researchers
are
iterating
more
on
the
data
and
need
the
tooling
and
hosting
to
be
able
to
do
that
well
and
preferably
in
a
very
reproducible
way.
So
that
way,
someone
can
just
pick
up
what
they've
done,
take
it
and
run
it
someplace
else,
and
so
this
is
active.
It's
like
a
little
bit
of
a
forcing
function
and
we
saw
this
initially
sort
of
with
OpenStack,
as
the
flexibility
like
opened
up
sort
of
a
lot
of
doors,
but
it
also
left
a
lot
to
be
desired.
C
You
still
had
to
sort
of
like
manage
VMs
and
then
manage
the
application
stacks
on
top
of
it
and
kubernetes
as
a
platform
solves
many
of
those
pain
points.
Now,
that's
not
to
say
it's
necessarily
easy
going.
There
are
certainly
some
challenges
so,
first
and
foremost,
many
institutions,
us
included,
have
investors
or
heavily
and
obviously
the
classic
model,
a
sequela
user
POSIX
infrastructure.
So
this
is
things
like
distributed:
shared
storage,
so
things
like
Joe
lustre,
gpfs
and
things
like
it's.
C
It's
kind
of
hard
to
take
a
api
at
and
you're
like
Oh
auth
to
identity
and
translate
that
to
like
a
UID
and
GID
that
can
propagate
down
to
the
file
system.
Well,
there's
some
workarounds
that
we've
done
both
myself
and
my
coworker
Jeff
was
over.
There
have
written
some
things
to
help,
take
care
of
that,
but
I
wouldn't
call
it
like
prod,
ready
in
any
way
shape
or
form
well,
I'll,
say
publicly
consumable
any
way,
shape
or
form.
C
The
another
thing
is
like
the
installation.
Experience
of
bare
metal
is
still
kind
of
a
pain
point,
but
there's
several
efforts
underway,
specifically
like
around
the
cluster
API.
That
might
help
improve
that.
The
other
thing
is
that
for
a
lot
of
our
researchers
like
slurm
and
torque,
aren't
like
that
they're
they're,
used
to
using
those
and
they're
honestly
had
a
bit
more
of
an
advantage
when
it
comes
to
scheduling
because
again,
like
a
classic
HPC
job
sort
of
like
falls
into
a
queue
and
has
a
wall
time.
C
C
Despite
all
those
challenges,
our
adoption
rates
have
been
good,
so
I'd
say
this
is
about
where
we
are
today.
Amongst
our
users,
general
users.
You
know
they
just
want
like
a
consumable.
Endpoint
like
they
put
in
a
request
for
a
Jupiter
hub,
we
say,
will
provision
them
Jupiter
hub
and
they're
happy
to
go
to
URL.
We
also
have
a
set
of
like
intermediate
users
that
don't
necessarily
want
access
to
kubernetes,
but
they're
just
happy
being
able
to
manage
it
all
through
get
so
they
want
to
be
able
to
push
their
own
image.
C
Make
a
few
changes
to
a
deployment.
They
don't
need
at
direct
access
to
the
cluster.
They
just
need
to
be
able
to
again
majid
through
get,
and
if
a
mistake
happens,
they
can
revert,
and
then
we
also
have
a
set
of
advanced
users.
Sort
of
like
that
group
I
was
describing
earlier.
I
was
building
like
a
whole,
app
stack
that
actually
just
want
to
you
sort
of
consume
kubernetes
itself
and
we're
able
to
give
them
that
access
by
you
know
essentially
giving
their
group
and
name
space
and
locking
it
down
with
our
back-end
PSPs.
C
So,
first
and
foremost,
we
plan
to
move
to
bare
metal.
One
of
the
big
things
has
been.
You
know:
hey
I,
have
these
27,000
cores
over
here
with
no
virtualization
layer
in
between
why?
Why
should
I
run
over
here
and
we're
hoping
that
you
know
that
that
might
placate
some
of
those
concerns?
The
other
thing
is
that
we
are
potentially
looking
at
other
platforms
or
other
architectures
I'm,
sorry,
so,
possibly
kubernetes
on
both
our
arm
and
power
infrastructure.
C
We
are,
as
I
sort
of
hinted
earlier,
working
on
like
active,
actively
improving
the
whole
like
UID
GID
mapping
problem.
So
they
could.
They
can
like
integrate
with
our
home
directories
as
well
as
automating,
some
of
the
like
Kerberos
and
candidate,
from
containers
to
more
easily
integrate
with
our
other
servlet
services.
We're
also
keeping
a
very
close
eye
on
some
of
the
hybrid
work
being
done.
C
So
Sai
Labs
has
a
slurm
operator
that
sort
of
lets
you
lets
us
integrate
with
our
sort
of
larger
HPC
cluster,
where
I
can
create
like
slurm
job
in
kubernetes
and
I
can
interact
with
that
in
normal,
like
kubernetes
way
as
well
as
any
other
tooling,
and
that
can
then
go
off
to
sort
of
our
HP
Club.
Alright,
our
HPC
cluster
do
its
thing
and
the
information
can
bubble
back
up
into
kubernetes.
C
Options
like
fair
share
and
I'm
honestly,
not
as
familiar
with
as
I'd
like
to
be
it's
a
pretty
new
project,
but
I
definitely
be
exploring
it
more
and
there's
also
one
area
that
I'm
going
to
be
diving
into
big-time
and
it's
outreach
and
training
for
both
our
faculty
and
students.
We
know
ku
Bernays
has
a
high
learning
curve
and
it
can
put
a
lot
of
people
like
off
initially
and
we've
sort
of
started
this
project
already,
we've
been
running
quite
like
sort
of
like
classes
internally
on
like
how
to
build
containers.
Well,
that's
also.
C
A
big
problem
is
like
in
the
research
world
you
find
like
containers
like
15
gigs
in
size
and
once
they
sort
of
go
beyond
that,
we
then
sort
of
teach
them
sort
of
how
to
use
kubernetes
and,
as
we
sort
of
like,
go
through
the
process
of
teaching
them.
They
sort
of
have
this
like
a
little
aha
moment
and
sort
of
start
to
realise
the
power
and
flexibility
of
that
offers,
and
so
like
with
these
sort
of
like
efforts
that
we're
doing
we're
sort
of
expecting
our
changes
to
go
from
this.
C
C
We
want
to
build,
offer
and
run
a
platform
that
enables
our
researchers
to
more
effectively
share
collaborate
and
honestly
do
their
research
and
Coronas
is
a
platform
that
lets
us
do
that,
and
while
there
are
definitely
some
rough
edges,
we
expect
those
to
improve,
especially
with
greater
adoption
and
with
that
I
will
hand
it
off
to
you
Lindsay,
to
bring
it
home.
Okay,.
B
And
we
did
want
to
leave
a
couple
minutes
for
questions,
but
I'm
just
gonna
do
a
quick
conclusion.
So,
to
recap
the
major
points
we
have
seen
that
kubernetes
provides
a
lot
of
great
things
for
researchers
as
well
as
research
providers.
There
are
lots
of
existence
of
existing
tools
that
facilitate
research,
workflows
such
as
argo,
cube
flow
and
cube
batch
and
there's
a
growing
demand
from
researchers
for
a
tool
that
does
everything.
Kubernetes
can
do
so.
B
We
do
have
a
quick
couple
of
suggestions
for
increasing
kubernetes
adoption,
most
of
the
provider
ones
bob
has
already
talked
about,
so
one
is
offer
kubernetes
for
people
to
consume.
This
is
kind
of
a.
If
you
build
it,
they
will
come
situation,
get
involved
with
the
kubernetes
community.
There's
a
lot
of
really
helpful,
knowledgeable
people
and
a
lot
of
them
will
be
more
than
willing
to
answer
any
questions
that
you
have.
B
If
you
join
a
cig
call
and
tell
them
your
use
cases
that
is
often
very
helpful
and
informative
for
developers
learn
as
much
as
you
can
again.
This
goes
in
part
with
the
previous
one
and,
as
Bob
mentioned,
provide
outreach
to
researchers
and
anyone
that
might
need
to
help
being
ramped
up
and
for
researchers
engage
with
the
research
institutions
if
you're
interested
in
running
kubernetes,
let
your
research
institution
know
and
tell
them
about
all
of
the
great
things
that
it
can
do
if
they're
not
aware
already
get
involved
with
the
kubernetes
community.
B
This
was
a
the
same
suggestion
for
providers,
learn
as
much
as
you
can.
This
might
also
sound
familiar
and
share
your
findings
widely
write.
A
blog
post
make
a
YouTube
video
make
do
a
proposal
to
speak
at
cube
con
whatever
helps
to
get
the
information
out
there
to
other
researchers,
and
that's
it
for
us.
So
there's
some
useful
links
that
you
might
be
interested
in.
C
B
D
A
C
If
you
go
to
github.com
slash
mr.
bobby
tables,
slash
Cates,
intro
tutorial,
it's
pretty
much
that
it's
also
in
the
like.
If
you
go
to
like
the
CN
CF
presentations,
repo,
it's
linked
under,
there
is
training.
We
also
our
container
training
up
there
and
then
a
lot
of
the
other
stuff
is
a
bit
more
one-on-one
or
with
smaller
groups.
It's
more
focused
on
like
improving
their
containers,
so
they'll
come
to
us
with
a
very
large
container
and
we'll
walk
them
through.
C
C
So
we
have
three
clusters:
we
have
one
that
is
our
general
services
cluster
and
the
sort
of
provide
services
to
our
other
clusters.
So
like
our
identity
management
and
things
like
that,
we
have
one
cluster
that
is
sort
of
the
public
facing
cluster
that
we
give
most
of
our
general
researchers
access
to.
And
then
we
have
a
third
cluster,
that
is
for
restricted
data
services
and
that's
very
tightly
locked
down
our
general
cluster.
Isn't
too
big?
It's
like
collectively,
I
think
512
gigs
are
Ram
and
like.