►
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
Panel: Control Plane vs Data Plane: Untangling the Tenets of Multitenancy - Erica Von Buelow, Red Hat; Sanjeev Rampal, Cisco; Ryan Bezdicek, Cray Inc.; Adrian Ludwin, Google; & Fei Guo, Alibaba
Virtually every organization over a certain size wants to be able to share their clusters between different sets of users. As a result, the Multi-tenancy Working Group is seeing increasingly high demand for higher-level features to support Kubernetes multi-tenancy. Unfortunately, each organization has different and often unspoken assumptions about what tenancy means to them, so different use cases and needs often get conflated. In this discussion, our panelists will share their proposals for the principles of multi-tenancy, according to both the type of concerns (control plane vs data plane) as well as the type of tenants (such as dev teams, production teams and third-party users).
https://sched.co/UaXF
A
For
everyone
hope
you're
having
a
good
start
to
your
cube
con.
So
we
have
a
interesting
panel,
a
great
panel,
actually
on
an
interesting
topic,
and
it's
interesting
because
just
a
week
ago,
survey
came
out
and
the
the
question
was
what
areas
of
Cuban
it
is.
You
think
the
community
needs
to
give
more
attention
to
in
2020,
and
this
was
one
of
the
top
topics.
So
clearly
this
is
on
top
of
many
of
your
minds
and
to
help
resolve
your
questions.
C
A
F
Scott
apologies,
I,
didn't
know
it's
quite
that
far
I
was
gonna,
make
Basecamp
halfway,
because
this
so
apologies
for
being
late.
I
am
a
Scotts
row
which
I
work
for
HSBC
Bank,
where
they're
a
global
bank
world's
largest
global
bank
and
I
am
actually
one
of
the
leads
for
the
internal
kubernetes
platform
for
the
design
that
would
begin
selling
globally.
A
And
myself,
I'm
Sanjeev
Rampal
I
am
also
involved
in
the
communities
working
group,
multi-tenancy
working
group
and
I'm
at
Cisco,
where
we
are
working
on
Enterprise
multi
cloud
platforms
using
technologies
such
as
given
it
is
so
we
will
start
off
with
an
audience
pool.
Perhaps
how
many
of
you
consider
that
you
would
be
decision-makers
on
Cuban
ideas,
clusters
in
any
form,
perhaps
as
operations
folks
or
a
technology
selection,
Evangel
technology
selection
decision
makers?
So
how
many
of
you
consider
making
decisions
about
Cuban
ideas
clusters
to
be
part
of
your
job?
A
Would
you
maybe
raise
your
hands?
Please?
Okay,
so
that
appears
to
be.
You
know
a
good
75
percent
of
the
room
now
off
those
how
many
of
you
feel
like
the
community
needs
to
do
more
in
the
area
of
multi-tenancy,
almost
the
entire
site.
So
this
validates
the
survey
results
as
well
that
this
is
an
area
the
community
needs
to
do
more
about
that's
great,
to
hear,
and
hopefully
a
panel
will
help
us
address
some
of
these
topics.
We
will
get
back
to
the
audience
with
more
questions,
because
we
want
to
make
this
interactive.
A
We
want
to
make
sure
that
the
panel
is
helping
the
larger
community
and
we
have
this
feedback
loop
to
help
us
come
up
with
the
right
sets
of
solutions.
So
let
me
start
with
the
fact
that
core
communities
was
never
quite
designed
with
multi-tenancy
to
begin
with,
there
are
no
users
and
projects
and
tenants
in
core
communities.
So
question
to
the
panel
is
was:
was
that
intentional
for
kubernetes
to
leave
out
multi-tenancy?
A
C
Well,
I
think
multi-tenancy
is
a
very
hard
problem
and
when
you're
launching
a
tool
like
kubernetes,
you
kind
of
optimize
to
solve
certain
problems
right
out
of
the
gate,
and
then
you
want
to
see
what
people
need
out
of
something
like
multi-tenancy
and
so
I
think
that
that's
prompt.
That
is
why,
in
the
initial
product
you
know,
capabilities
of
kubernetes
as
a
platform
multi-tenancy
was
left
as
a
problem
to
be
solved
based
on
feedback
as
people
adopted
and
started
to
use.
It
sure.
D
If
I
can
I
completely
agree
with
that
and
I
think
that
some
of
the
things
we've
seen
as
a
part
of
the
multi
tendency
working
group
have
validated
that
decision,
because
when
we
go
out
to
the
community-
and
we
say
what
is
it
that
you
need
for
multi-tenancy-
we
get
wildly
diverging
expectations
about
what
multi-tenancy
is
and
how
it
should
be
solved.
And
so
I
think
waiting
for
these
patterns
to
appear
and
emerge
has
has
been
the
right
idea.
E
Yeah
I
just
wanted
to
echo
what
Tasha
was
mentioning.
I
think
this
is
Martinez.
It's
a
tough
problem
to
solve
it's
not
very
easy,
I
mean
if
we
build
a
platform,
you
must
start
from
simple
everybody
with
you
this
way
so,
but
once
you're
reaching
a
certain
stage
once
to
your
commercial
age,
you
want
to
production,
you
have
more
as
well
heat
ammonia
fancy
program,
but
my
personal
feeling
is
about
the
communities
is,
although
you
know,
the
coke
API
server
is
not
designed
for
multi-tenancy,
but
raised
in
hell.
A
I
also
want
to
clarify
that
the
topic
described
our
panel
as
control
plane
versus
data
plane,
but
really
we're
talking
about
all
aspects
of
multi-tenancy
here.
Essentially
there
are
control
plane
aspects
like
having
explicit
notions
of
tenants
and
projects
and
data
plane.
Isolation,
technologies,
like
you
know,
SEC,
comp
or
cata
containers
and
so
on.
So
the
working
group
and
the
community
in
general
is
sort
of
addressing
it
from
a
complete
control.
Plane
and
data
plane
perspective.
So
this
is
really
not
about
control,
plane
versus
dinner
plane,
but
more
like
control,
plane,
plus
data
plane.
A
Let's
see
so
maybe
we
can
go
around
and
ask
what
each
of
you
feels
is
your
definition
of
multi-tenancy
and
how
you
might
differentiate
between
soft
multi-tenancy
and
hard
multi-tenancy.
So
maybe,
let's
get
started
with
Scott.
What
would
you
define
multi-tenancy
to
be
and
do
you
differentiate
between
soft
and
hard
multi-tenancy
yeah.
F
F
So
in
a
multi-tenant
world,
we're
gonna
have
our
say:
40
node
cluster,
that's
going
to
have
multiple
businesses,
multiple
apps
and
it
could
be
multiple
tiers
of
apps,
so
it
could
be
a
2-0
tier
1
tier
2,
app,
that's
actually
still
evolving
internally,
but
we
will
have
different
businesses
on
one
cluster
and
different
types
of
APIs
load
so
getting
into
the
second
part,
which
would
be
the
soft
versus
a
heart
multi-tenant.
It's
a
conversation.
That's
ever
never
stops
internally,
because
there
are
no
true
good
definitions
of
it.
F
Honestly
I
put
everything
almost
in
the
soft
multi-tenancy,
so
we
use
a
lot
of
software
for
an
example
to
do
all
that
and
we'll
cover
that
a
little
bit
later.
I
think
you
know
want
to
be
do
that
now
right,
so
so
that
in
the
hard
multi-tenancy
we
were
thinking
more
individual
clusters.
That
might
be
fronted
with
more
security,
but
for
the
most
part
we
go
into
the
soft
I'm.
A
Hearing
a
bit
of
a
boom
on
the
speakers
is
the
sound
okay
for
everyone,
all
good
okay.
Maybe
it's
good
jamming
us
to
the
thing
so,
and
the
other
thing
is:
god
is
that
you're
in
the
financial
services
industry,
where
even
internal
tenants
are
needing
to
be
isolated,
almost
like
external
tenants
right?
So
could
you
speak
to
the
fact
that
in
your
specific
industry,
there's
no
such
thing
as
soft
multi-tenancy?
It's
really!
You
know
your
separation
between
your
intra
enterprise
chances,
almost
as
good
as
external
attendance.
Any
any
comments
on
absolve.
F
So
obviously,
we're
highly
regulated
audited
all
the
time
regulators
coming
in
so
you're
right.
So
typically,
even
when
I
talked
about
the
multi
tenants
at
a
tier
0
tier
1,
through
2
application,
we
treat
them
all
the
same,
meaning
yeah.
We
we
use
PSPs,
we
use
network
security
policies,
we
don't
necessarily
trust
just
PSP,
so
we
have
opa
on
top
of
that,
just
because
we
do
give
some
elevated
permissions
to
the
namespace
and
somebody
could
potentially
create
a
PSP
that
overrode
what
we
send
down.
So
we
use
OPA
to
catch
that.
B
For
me
right
so
multi-tenancy
in
general
is
kind
of
a
lot
of
term
because
to
most
people,
there's
very
strict
meaning
of
multi-tenancy
expression
from
an
auditing
perspective.
So
when
we
say
soft,
multi-tenancy
mode,
people
kind
of
you
know
scoff
at
that.
But
for
me
Sabaton
teen,
it's
easy
right!
It's
it's
people
acting
in
you
know
good
faith
really
of
sharing
shared
resources,
but
not
wanting
to
accidentally
or
even
purposely.
B
You
know,
destroy
them,
for
you
know:
takedown,
yeah,
networking,
storage,
etc,
and
so
my
hard,
my
presidency
is
really
where
you
there's
no
trust
and
you're.
Trying
to
you
know
compromise
where
you
can
and
I
think
you
know
for
right
now.
Corbin
is
just
starting
to
get
to
the
point
where
hard
maps
and
seas
may
be
viable
with
some
of
the
new
security
or
on
container
and
hardware
sharing
so
Tasha.
B
C
So
from
my
perspective,
soft
multi-tenancy
is
interesting
because
it's
something
we
think
we
could
achieve
today
with
the
features
and
capabilities
in
kubernetes
and
when
I
think
of
what
that
definition
is.
It
means
that
you
do
have
this
shared
trust
that
perhaps
it's
a
internal
to
your
business,
shared
kubernetes
control,
plane
people
have
some
sort
of
incentive
to
not
DDoS
each
other
and
to
respect
their
resource
limits.
But
you
don't
have
to
fundamentally
architect
every
layer
of
your
stack
to
defend
against
that
and
then
with
hard
multi-tenancy.
You
do.
D
So
in
those
cases
that
will
look
very
different
from
multi-team
tendency
in
your
WordPress
users
will
not
have
access
to
cube
cuddle.
They
won't
have
access
to
the
control
plane,
but
you
might-
and
you
might
also-
and
and
so
there's
going
to
be
some
kind
of
gradation
between
amount
of
access
that
the
customers
have
to
the
cluster
from
zero
to
full
the
amount
that
you
trust,
the
workloads
that
are
running
on
them
and
the
same
is
true
of
the
spectrum
from
hard
to
soft.
It's
not,
of
course,
binary.
D
E
I,
just
I
just
wanted
to
I,
don't
know.
Maybe
the
point
is
oh,
my
my
point
is
I
think,
especially
in
our
community's
context.
The
Mauri
tendencies.
As
the
point
is
everybody
sharing
the
same
the
resource.
So
that's
a
point.
So
I
see
a
lot
of
some
solution:
core
multi-tenancy
each
put
each
user
has
dedicated
cluster.
Yes,
that's
Marty
Hennessey,
so
so
so
overall
I
think
my
point
is
especially
my
experience
of
using
kubernetes,
because
Alibaba
has
been
using
community
for
a
long
time.
Our
biggest
challenge
is
always
the
isolation.
E
The
problem
may
already
been
largely
solved
by
VM
kind
of
word,
but
not
in
a
content
network.
So
the
kubernetes
is
great
for
our
application,
but
I
would
say
it's
not
very
friendly
to
the
ops.
So
we
see
a
lot
of
problem
that
introduced
by
the
isolation.
So
that's
why
I
think
we
should,
although
I
think
the
Marty
tendency
the
working
group,
whatever
so
I
I,
think
I'll
prime
many
of
our
solution,
a
focus
on
the
isolation
perspective.
E
A
You
so
is
it
also
fair
to
conclude
that
as
people
start
running,
containers
more
on
bare
metal,
the
need
for
multi-tenancy
will
become
more
important
because
you
really
can't
give
everybody
their
own
bare
metal
clusters.
The
way
you
can
give
them
their
virtual
machine-based
clusters.
So
any
thoughts
on
this
being
related
to
the
move,
to
having
more
containers
run
on
bare
metal,
as
maybe
Scott,
are
you
running
Kuban
Edison,
docker
containers
on
bare
metal,
or
is
that
your
plan,
or
how
do
you
see
that
it's.
F
An
interesting
question
we
actually
are
on
our
virtual
machines,
because
we
are
a
big
Google
customer
and
we
do
run
G
Kem
pram,
which
is
a
requirement
to
run
vSphere
right
now.
However,
we
have
been
working
with
Google
to
potentially
move
that
to
bare
metal,
which
has
been
our
preferred
platform
for
multiple
reasons,
but
we
do
want
to
go
to
the
bare
metal
rather
than
a
virtual
side.
But
right
now
we're
pure
virtual
right
is.
F
F
B
So
we
actually
our
bare
metal
right.
So
for
us
it
is
important,
because
when
we
can't
do
multiple
clusters,
mainly
because
we're
telling
our
customers
they
need
to
buy
bare
metal
forever
kubernetes
cluster,
they
need
so
for
us.
Multi-Tenancy
at
the
bare
metal
level
is
very
important,
which
is
why
we
were
involved
in
the
working
group
trying
to
achieve
that,
and
it
is
quite
difficult
today,
nearly
impossible.
Anybody.
C
Yeah,
so
what
we've
been
working
on
is
making
containers,
have
the
isolation
and
security
premises
of
virtual
machines
and
then
have
the
performance
of
bare
metal
with
some
of
the
work
that
we're
doing
with
project
specific
with
hypervisor
optimized
pods,
and
so
from
that
perspective,
we're
not
concerned
with
bare
metal.
In
that
regard,
we
want
to
understand
why
people
think
they
need
bare
metal
and
then
see.
If
there
are
other
things
we
can
do
to
provide
them
a
really
great
performance
experience
that
still
allows
them
to
keep
the
control
plane
for
their
virtualized
infrastructure.
E
So
you
know
Alibaba,
actually,
the
entire
ecommerce
applications
very
big
amount
of
applications
run
in
bare
metal
machines
for
a
long
time
so,
but
you
can
treat
that
as
a
singleton
in
the
world
so
not
very
fit,
but
for
Marty
talents
I
think
we
have
a
product
which
is
running
on
bare
metal
on
top
of
the
kubernetes
is
multi-tenancy
solution
that
the
key
point
again
as
I
mentioned
before
is
isolation.
So,
besides
their
normal
things,
that
I
can
think
of.
We
use
the
color
container,
like
Samba
song
time
for
the
countenance.
D
This
also
yeah
go
ahead.
Aging,
I
I
was
just
gonna,
say
I'm,
obviously
on
gke,
you
are
running
on
virtual
machines
and
not
on
bare
metal,
but
there
are-
and
certainly
if
you
are
running
on
bare
metal,
you
will
want
to
have
as
few
clusters
as
you
want.
But
even
if
you
are
running
virtual
eyes,
it's
not
free,
obviously
to
have
lots
of
clusters
because
there's
a
management
overhead
there's
a
complexity
overhead,
they
all
have
to
be
upgraded
and
maintained
in
some
way.
This
sounds.
A
Like
a
question
which
would
be
useful
to
get
the
audience
feedback,
how
many
of
you
feel
like
your
end
goal
is
to
run
containers
on
bare
metal
and
how
many
you
feel
like
you
can
live
with
running
containers,
virtual
machines
forever?
Okay.
So
this
is
an
interesting
data
point
that
a
lot
of
you
feel
like
to
be
just
to
summarise
to
everyone.
A
It
sounded
like
about
more
than
half
the
folks
of
the
over
here
feel
like
they
want
to
continue
running
containers
on
virtual
machines,
and
maybe
a
good
25
to
30
percent
expect
to
be
running
on
bare
metal.
So
that's
an
interesting
data
point
that
maybe
we
will
have
our
panel
express
their
thoughts
about.
A
Let's
see
so,
let's
talk
a
little
bit
about
the
working
group,
so
a
lot
of
you
are
in
the
kubernetes
working
group
for
multi-tenancy.
Could
you
maybe
summarize?
What
are
the
activities
and
projects
that
are
being
undertaken
in
this
working
group?
For
those
of
you
that
might
not
be
familiar?
There
is
a
working
group.
A
We
are
loosely
affiliated
with
the
Sagat,
but
it's
an
independent
team
and
most
of
the
folks
on
the
panel
here
are
actually
participating
actively
and
that
so
maybe
they
can
tell
us
what
the
group
is
doing
and
you
know
how
this
solves
problem
so
I
think
Ryan
and
others
Scott.
You
may
not
be
directly
familiar
so
start
with
Ryan
go
ahead,
Ryan
all.
B
C
C
We
have
a
couple
different
incubation
projects
which
Adrienne
and
Faye
lead,
and
we
have
so
they'll
both
talk
about
those
in
a
little
more
detail,
I
expect.
But
we
also
have
some
other
cool
things
that
people
really
like.
So
we
went
to
everyone
who
is
involved
in
our
working
group
and
had
them
tell
us
why
they
wanted
multi-tenancy
what
they
wanted
out
of
multi-tenancy
and
what
they
were
seeking
to
achieve.
C
And
then
we
have
a
doc
that
has
that
that
a
bunch
of
other
working
groups
and
SIG's
have
found
very
valuable
as
they
try
to
understand
what
features
they
need
to
bake
into
the
work.
They're
doing
so,
if
you
have
a
use
case,
we
would
love
for
you
to
just
pull
up
that
Google
Doc
and
contribute
your
use
case
so
that
we're
considering
it
as
we
build
out
our
tools.
C
D
I'll
talk
very
briefly
about
the
project
that
I'm,
leading
and
I
should
emphasize
that
the
the
whole
group
helps
out,
especially
ryan,
has
been.
Everybody
on
this
group
has
been
fantastic
since
we
started
working
on
this
project,
so
I'm
working
on
hierarchical,
namespaces,
the
HNC,
hierarchical
namespace
controller.
This
is
a
project
to,
as
the
name
implies,
add
hierarchy
to
namespaces,
which
what
this
means
is
that,
if
you
create
policy
objects
in
parent
namespaces,
they
can
be
propagated
to
the
children
and
all
of
the
descendants.
So
it's
a
good
way
to
enforce
policy
across
an
organization.
E
Yes,
I'm
going
to
briefly
describe
the
project
I'm
leading
the
working
group.
The
name
is
water
cluster.
So,
very
briefly,
the
idea
is
we
assign
each
tenant
a
dedicated,
API
server
so
to
the
older
tenant,
we'll
share
the
same
super
API
server
which
provide
the
party
resource.
So
the
core
idea
is
to
give
a
strong
isolation
between
tenants,
especially
in
a
control
plane
in
terms
of
working
in
the
working
group.
I
think
our
walking
group
is
pretty
much
is
a
use
case
driven
so
so
just
feel
free
to.
You
know
express
your
use
cases.
E
What
exactly
you?
What
in
Tallinn
see
solution,
chimera
solution
looks
like
in
your
or
whatever
and
I
think
we
will
just
coordinated
with
the
sauce
and
give
the
design
to
you.
We
can
you
can
change
your
colony
is,
is
not
fixing
yeah,
so
that's
a
whole
reason
that
I've,
we
I
think
we
already
cracked
some
feedbacks
and
we
definitely
one
more
so
yeah.
C
C
A
Scott,
let
me
turn
to
you,
tell
us
what
you're
doing
in
your
production
environment.
How
are
you
implementing
kubernetes
multi-tenancy
and
if
I
might
throw
in
a
third
question
in
that
little
mix?
What
would
you
like
the
community
in
the
working
group
to
do
for
customers
like
yourself
to
help
you
about
a
solution?
But
let's
start
with
what
you're
doing
today,
yeah.
F
I
mean
some
of
the
projects
that
you
guys
have
working
like
the
hierarchical
namespace
is
interesting
to
me,
because
we
use,
of
course,
anthos
config
management,
so
very
similar
just
implement
it
differently.
So
we
put
our
if
anybody's,
seen
ACM
we
put
very
similar
policies
in
github
that
can
create
the
namespace.
It
also
will
create
our
PS
PS
and
basically
anything
else
you
want
to
put
in
there
to
TMO
file.
So
what
we
do
is
every
namespace
is
labeled
and
annotated,
like
you
were
saying
as
well,
using
what
we
call
HSBC
metadata.
F
So
I
can
tell
you
what
business
it
is,
their
queues,
their
maintenance
windows,
so
we
can
link
those
and
we
do
use
those
for
other
policies
as
well,
and
you
have
to
have
those
in
place
to
begin
with
or
you've
got
major
problems
later.
So
that's
one
of
our
that
was
one
of
our
biggest
design.
Factors
is.
What
are
we
going
to
use
for
this,
and
to
many
is
also
a
bad
thing.
So
come
up
with
the
right
information,
and
hopefully
we
pull
it
well.
F
In
our
case,
we
pull
from
a
database
as
a
source
of
truth.
We
don't
let
you
enter
free
form
fields.
So
when
you
request
a
namespace
which
is
all
self-service
using
one
hour-
oh
I,
DC
provider,
you
select
all
this
information
creates
the
namespace
for
you
and
it
will
then
create
your
access
role
and
that
depends
on
what
you're
requesting.
So
you
might
have
admin
for
the
namespace.
You
might
have
viewer
only
privileged
accounts
get
admin.
Non-Privileged
are
gonna
use
read-only.
So
that's
just
the
first
part,
as
we
mentioned
earlier.
F
We
of
course
use
the
PSP's
to
lock
things
down
a
host
path.
You
don't
want
somebody
suddenly
remounting
root
on
the
host
and
grabbing
every
file
on
it,
but
at
the
same
time,
since
we
do
give
you
a
lot
of
control
in
the
namespace,
we
use
open
policy
agent
as
well,
so
using
OPA.
We
will
check
for
the
same
information.
F
We
also,
of
course,
use
network
policies
by
default,
meaning
you
get.
No
egress,
no
ingress,
you
better
do
something,
or
else
just
gonna
sit
there
and
talk
to
itself.
That's
been
a
little
bit
of
attic
because
not
everybody
knows
to
create
that,
so
it
does
generate
phone
calls.
It's
a
great
idea.
We
document
it,
but
people
still
have
questions
on
top
of
that.
We
because
we
do
use
gke.
Our
control
plane
is
isolated
from
the
user.
F
So
we
have
the
admin
control
plan
and
you
got
the
user
control
plane,
so
it's
actually
great
design
for
gke.
So
all
those
together
have
been
very
good
for
us
and
to
stop
the
bad
neighbors
I.
Think
like
you
were
mentioning
with
resources.
We
of
course
have
resource
quotas
on
everything,
soft
quotas
and
hard
quotas,
and
you
can
request
an
increase
to
that
through
a
portal,
but
by
default
we
have
certain
rules
set
up
for
this.
F
Many
CPUs
PBC's
as
many
pods
just
to
stop
people
from
just
consuming
everything,
and
there
has
to
be
some
justification
to
that
and,
of
course,
there's
an
internal
chargeback
mechanism.
For
all
of
that,
we
of
course
have
logging
in
place
and
reporting,
so
I
know
how
to
scale
my
clusters
at
60%,
I,
better
start
buying
something
or
starting
to
expand
and
scale
out
and
at
80%
I
should
have
that
in
place
or
I'm
in
big
trouble
going
forward.
F
And
that's
that's
great
cuz
I
know
we
got
introduced
a
couple
months
ago,
so
it's
great
that
I
can
give
some
feedback
into
this.
The
projects
on
the
incubator
are
great,
so
the
virtual
cluster
I
did
ask
my
Google
contacts
about
the
hierarchical,
namespace
and
ACM.
So
I
know
that
probably
virtually
at
some
point,
so
the
projects
that
are
coming
really
help
the
only
other
one
I've
seen
and
I.
Had
this
conversation
with
some
storage
vendors
today
is
something
with
PVCs
to
help
with
security
I'm.
A
Mean
say
what
more
do
customers
need,
so
just
a
little
semantic
thing,
but
yeah
point
taken
I
think
so
it's
a
great
example
from
Scott
here
that
you
can
do
a
lot
with
namespaces
today,
but
there's
a
lot
of
tribal
knowledge.
There's
nothing
standardized!
You
know
you
read
this
blog
and
you
read
that
blog
and
you
read
apply
the
PSP
and
then
am
I
doing
the
exactly
the
same
level
of
isolation
that
somebody
else
is
doing.
A
I
don't
know
right.
So
one
of
the
goals
here
is
also
in
the
community
is
also
to
have
develop
a
set
of
benchmark
Suites
so
that
you
can
run
your
clock,
run
these
Suites
on
your
cluster
and
say:
ok,
my
cluster
conforms
to
benchmark
profile.
A
so
I
know
if
I
take
it
to
another
cluster
which
benchmark
profile
be
it'll,
have
the
same
behavior.
A
Otherwise,
if
if
multi-tenancy
is
based
on
ad-hoc
tribal
knowledge,
I
don't
know
whether
I'll
get
the
same
isolation
when
I
move
my
application
from
cluster
a
to
cluster
B,
because
it's
all
based
on
whatever
cluster
admin
decided,
was
adequate
for
multi-tenancy.
So
that's
a
that's
another
thing
here.
Actually,
let's,
let's
take
another
audience
poll
here,
which
is
we
talked
about
control,
plane
versus
data
plane.
What
do
you
all
feel
is
needs
more
solutions
from
the
community.
You
feel
like.
We
need
more
control,
plane
solutions
like
creating
tenants
managing
tenants.
A
You
know
the
the
control
plane
for
the
virtual
hierarchical
namespace
or
do
you
feel
like
you,
need
more
data
plane
solutions
like
recommendations
on
using
various
container
sandbox
in
technologies?
So
the
question
is:
what
do
you
think
control
produces
database,
so
how
many
of
you
feel
like
you
need?
The
community
needs
more
work
in
the
area
of
control
for
multi-tenancy,
so
it
sounds
like
about
a
quarter
of
you.
How
many
I
feel
like
you
need
more
the
community.
There
needs
to
do
more
in
the
data
plane
area
to
support
multi-tenancy
a
little
bit
more.
A
So
it
sounds
like
there's
a
lot
of
projects
that
are
going
on,
and
maybe
let's
get
the
opinion
of
the
panels
so
panel.
Your
thoughts
on,
let's
say
data
plane,
isolation,
technologies
have
you
used
set
comp
or
Kara
containers
or
fryer
crack
orgy
visor.
Your
thoughts
on
your
and
your
experience
on
on
any
of
these
technologies,
which
you'd
like
to
share,
go
ahead
right
right.
B
So
well,
we're
still
investigating
right,
so
we're
here
to
discover
that
to
me,
firecracker
seems
quite
amazing
and
you
know
we're
looking
at
that
quite
heavily
gee
advisor
as
well.
So
those
are
you
know
what
we're
looking
at
closely,
but
we
haven't
decided
and
we're
hoping
others
can
you
give
us
a?
Is.
A
B
So
that's
part
of
it
right
is.
We
have
multiple
use.
Cases
outside
of
kubernetes
for
firecracker
could
help
us,
and
so
it's
kind
of
that's
one
of
the
nice
things
where
that
gives
it
to
us
and
that's
why
we're
looking
at
kind
of
a
broad
scope?
But
you
know
it
also
can
be
used
within
the
kubernetes
concept,
so
that
helps
its.
A
F
Cuz
and
being
gke
customers
g
vice
is
kind
of
a
given
that
we've
been
talking
about
that.
It's
still
not
you
know.
Quite
there.
Obviously
it's
run
everywhere.
We
of
course
from
out
the
host
operating
system.
We
do
go
into
se.
You
know
selinux,
second
comp
we're
trying
to
set
up
our
policies
for
that
today.
Actually,
we
just
don't
have
those
ready.
We
haven't
had
a
need
to
fortunately,
but
we
definitely
have
to
have
those
and
we're
gonna
make
that
a
standard
across
all
the
providers
and
also.
A
I
think
a
lot
of
these
technologies
are
also
in
alpha
state
at
kuna
on
the
kubernetes
api.
Some
of
them
will
transition
into
beta
and
I
think
we'll
see
more
traction
in
the
coming
few
months
on.
You
know,
with
with
more
runtime
classes
and
so
on.
You
know
etosha
any
thoughts
from
your
side
on
you're,
either
personal
experience
or
what
you've
heard
about
different
data
plane.
Isolation,
technologies,
yes,.
C
So
something
that
we
are
developing
for
vSphere
project
Pacific
is
what
we're
currently
calling
a
vSphere
native
pods,
which
is
a
container
runtime
optimized
for
running
on
the
hypervisor,
and
so
what
that
gives.
You
is
complete
isolation
of
your
various
pods
using
the
VM
boundary,
and
so,
if
you
break
out
of
your
container
there's
nowhere
else
to
go,
and
they
also
yeah
are
very
lightweight
and
fast.
So
that's
really
kind
of
caught
a
container
like
the
other
cool
thing
that
I
was
hearing
about
this
morning.
I
think
locky
did
a
keynote
on.
C
It
was
talking
about
the
new
features
that
are
coming
from
the
different
processors,
so
like
Intel
AMD
are
starting
to
provide.
Well
specifically,
this
was
Intel
starting
to
provide
container
isolation
link
down
to
the
memory.
So
if
someone
comes
in
and
changes
like
a
single
bit
of
your
memory,
like
the
entire,
the
entire
checksum
is
invalidated,
and
you
know
that
someone's
been
messing
around
and
so
from
that
perspective.
I
think
there's
gonna
be
some
really
nice
low-level
features
that
will
provide
significantly
better
isolation
for
containers.
I,
don't.
D
Have
a
lot
of
personal
experience
and
in
sandboxing
technologies.
What
I
do
know
about
gee
visor
is
that
it
has
been
run
in
production
for
a
very
long
time
and
it's
a
very
powerful
sandboxing
technology
and,
as
you
know,
it's
relatively
easy
to
use.
It's
in
beta
I
believe
in
NGK
right
now
and
I.
Don't
have
a
lot.
E
Yes,
I
mean
I,
never
even
any
I
think
we
have
both
use
case
for
kata
and
divisor,
but
primarily
our
use
cases
need
sandbox
container.
We
use
cotta
container.
So,
yes,
the
samples
container
can
greatly
help
the
isolation
problem,
but
the
issue
is
people
is
switching
from
wrong
C
to
cotta,
so
they
have
a
standard
for
the
performance.
So
if
you
switch
to
vmq
cotta
they're,
probably
fine,
but
this
we
switch
from
ronzi
to
cotta.
They
always
ask
about
the
performance
even
is
about
10%
slowdown.
E
A
A
C
F
And
honestly,
probably
just
since,
since
we
are
highly
regulated,
support
is
probably
the
biggest
thing
right
there.
We
need
to
prove
in
many
instances
that
we
have
some
kind
of
escalation
in
support
and,
if
I
say,
hey,
I
wrote
that
and
it's
not
gonna
be
acceptable
depending
on
the
application.
So
if
you
can't
prove
that
there's
hefty
fines
and
other
issues
that
could
happen
so
yeah
putting
it
putting
it
in
sight
does
help.
It
also
can
assist
us
in
the
fact
of
it's
part
of
the
base
product.
F
We
should
be
able
to
use
it
in
theory
across
all
providers.
Of
course,
we're
multi-cloud.
So
we're
not
just
internal
we're
hybrid
all
over
the
place
and
we've
run
into
that
today,
where
I'm
gonna
have
something
that
runs
inside,
because
I
can
but
I
can't
do
it
in
Amazon
I
can't
do
it,
man
usually
do
it
in
GCP,
just
because
we're
using
GTE.
It's
not
always
true,
though,
but
for
the
most
thing
so
that
that's
probably
the
main
reason.
D
I
would
add,
standardization
is
very
important
in
the
the
I.
Don't
know
who
was
at
the
GM
talk
earlier
today,
but
they
were
talking
about
how
oh
we
wish
cube.
Flo
supported
this
standard
model,
but
there
is
no
standard
model,
so
the
more
we
can
standardize
the
more
we
can
standardize
the
model,
which
is
one
of
the
reasons
why
hn
c
looks
like
ACM
like
Anthes
config
management,
the
more
we
can
standardize
the
model,
the
better
will
be
for
the
entire
ecosystem
and
better
for
portability.
You
can
move
applications.
B
G
Anyone
heads
out
I
would
like
to
remind
you
all
to
do
the
feedback
on
this
session.
If
you'd
like
to
see
more
panels
like
this,
we
really
want
to
hear
that
from
you.
If
you
got
something
out
of
this,
if
there
are
any
points
of
feedback
that
you'd
like
to
share
with
the
panelists
or
us
at
the
CN
CF,
please
do
I
think
we
can
take,
maybe
one
more
question
and
then
we'll
we'll
wrap
it
up.
So
there's
one
back
here.
H
You
know
so
my
questions,
quick,
it's
just
for
Scott.
It's
do
you
have
a
white
paper
or
anything
on
your
setup
sounds
like
you've
gone
pretty
far
into
this
and
have
some
pretty
nice
things
already
done
for
yourself
and
it'd,
be
nice
to
see
some
of
your
OPA
policies
as
well
as
like
some
of
like
how
you're
building
your
different
namespaces
and
things
like
that
ya.
F
Know
that's
funny.
Google
asked
us
for
the
OPA
stuff
too
saying
what
we
would
suggest.
So
what
we're
working
on
that
as
an
organization
you
can
imagine
to
certain
things
we
they
don't
like
us
publishing
talking
great.
We
can
definitely
talk
if
you
want
any
questions.
I've
got
no
rush
to
even
you
know,
give
you
my
email
feel
free.
My
kid
always
I've
been
debating
how
I
can
I'm
gonna,
say
water
down
what
I
have
to
make
it
more
generic,
so
take
out
the
word
HSBC
put
in
bank
or
something
but
yeah
yeah.