►
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
Access Policies for Hybrid Cloud Environments - Ruiyi Wang, Google
A typical admin/developer today needs to deal with many access policy systems: K8s RBAC, Istio RBAC and ABAC, OPA, policies from cloud vendors including IAM, firewall rules, quota etc. Reasoning about access across these systems is getting more and more complex and difficult. We want to invite those who build access policies and those who use access policies to come together to discuss the vision of a secure, easy to use and consistent/coherent access management experience across infrastructure, services and applications. And more importantly: how do we get there?
To learn more click here: https://sched.co/FuKy
A
Which,
which
means
I
do
very
little
talking,
would
do
a
lot
of
talking
and
I
see
some
people
who
don't
speak
Chinese.
Some
people
do
some
people
with
earpieces.
Some
don't
I
can
try
to
try
to
repeat
stuffing
in
both
Chinese
English.
So
that's
easy
for
everybody
Doge
anyway.
She
got
bird
of
feather
the
session.
So
it's
you
should
I
got
a
participation,
control
yeah.
What
can
I
do
to
go
and
film
on
the
YT
now
dango
Kaito
schwa
johani
120.
A
Now,
okay,
though
one
ahold
is
your
woman
cancel,
but
the
more
on
da
da
da
da
Chen
to
Lehtonen
participate
the
whole
you
know,
have
a
cover
discussion
together:
okay,
well,
the
to
hydro,
sure
I'm
gonna
spend
20%
I'm
just
setting
up
the
contact,
so
wait
a
leash,
I
wish
of
a
J.
She
goes
show
me
on
pylons
face
well,
my
way
to
me:
I
teach
you
the
province
base
and
then
80
percent
turf
is
gonna,
be
discussion,
IRA
first
and
then
I
think
it's
a
province.
Cartman,
that's
gonna,
go
be
young.
A
A
Okay,
so
the
first
thing
I
want
to
paint
a
picture
for
any
of
you
who
work
with
cloud,
especially
a
work
with
Mount,
multi-cloud
and
kubernetes
in
this.
Do
when
you
think
about
what
life
looks
like
for
use
or
an
admin
when
they
come
when
it
comes
to
dealing
with
access
policies?
This
is
what
I
deal
with
today.
So
it's
a
pretty
big
mess
on
the
infrastructure
layer.
They
can
use
one
or
more
clouds.
A
A
I
am
a
tan
yellow
today
are
of
these
application
layers
then
lead
to
your
data,
so
you
can
take
a
key
office
service
accounts
to
then
get
access
to
your
data,
which
can
also
be
accessed
by
I.
Am
so
if
you
try
to
answer
the
question
of
which
users
can
access
my
data
or
why
can
I
use
and
not
access
my
data?
It's
really
complicated
you're.
Looking
through
these
dozens
of
different
access
policies
to
try
to
figure
out,
you
know
what
you
need
to
change
or
when
you're
making
a
change.
A
What
is
the
total
effective
policy?
What
is
the
the
consequence
of
the
change
you're
making
so
and
then
what
we
notice
is
that
a
lot
of
users-
and
that
means
look
at
it,
so
there's
no
paralyzed
there.
They
go
I'm,
not
really
quite
sure
how
to
what's
the
best
practice
around
all
these
policies,
how
to
troubleshoot
them,
how
to
understand
them,
how
to
use
them
in
combination
with
each
other.
A
So
since
we've
got
more
than
a
dozen
of
policies
there
and
in
fact
what
we
listed
those
as
soon
as
show
that
you,
when
I
find
engineer,
he
said
no,
it's
not
a
complete
picture.
You
also
need
to
think
about
authentication,
binary
and
binary
authorization
quota.
Other
policies,
there's
no
entry
list
right.
A
A
We
have
to
learn
how
to
express
the
same
thing
in
many
different
ways
and
in
best
practices
right
most
of
the
cloud
customers
come
to
us
and
say:
hey,
don't
give
me,
you
know,
set
of
22
features
of
20
different
policies.
Tell
me
how
to
use
them.
Give
me
the
opinionated
best
practice.
Give
me
a
solution.
Give
me
a
template
that
I
can
start
from.
When
you
have
these
tiny
policies,
every
policymaker
has
their
own
best
practice.
A
They
come
up
with
their
own
recommended
defaults,
but
when
you
put
them
together
do
that
to
make
sense,
there's
nobody
doing
enter
and
testing
there's
nobody
thinking
through
the
effect
of
the
best
practices
when
they
work
in
conjunction
with
each
other.
Let's
pick
some
other
examples.
Error
message
is
another
radius
tricky
place
right,
there's
no
common
standard
for
producing
error
message
for
access
denied
right,
it
doesn't
tell
you.
Sometimes
it
just
says
access
denied.
It
doesn't
tell
you
it's
because
of
our
back.
A
It's
the
curse
of
sto
layer,
it's
because
of
networking
policy,
and
then
it's
you
know
actually
add
me
to
figure
out
on
my
own,
exactly
which,
which
policy
even
to
start
with
other
things
like
you
know,
things
like
explain,
ability
and
simulations
start
to
become
more
and
more
important.
The
more
layers
you
have
the
more
sophisticated
the
system
get
any
add
me
when
a
user
wants
to
figure
out
when
I'm
making
a
change.
A
What
is
the,
what
is
the
end
result
right,
how
it's
gonna
impact
my
production
traffic
and
when
something
goes
wrong,
how
do
I,
how
do
I
go
about
fixing
it?
What
are
different
options?
I
have
and
again,
if
every
one
of
these
policy
is
working
isolation
in
their
own
silo,
then
we
don't
have
an
entering
story
to
help
anybody
to
really
manage
access
control
for
the
hybrid
cloud.
A
So
you
do.
We
want
to
pick
anything
else.
Also.
You
know
labels
and
groups.
Some
policies
go
with
a
hierarchical
structure.
Right.
You
know
the
group
resources
together,
there's
a
kind
of
ownership
structure,
typically
in
AD
any
out
app
policies
inherited
them
a
strict
hierarchy
in
kubernetes.
You've
got
namespaces
which
are
very
flat,
and
then
we
can
put
labels
on
them.
We
use
label
selector,
that's
a
very
kind
of
a
backhand
flat
type
of
policy.
Again
when
you
put
a
hierarchical
policy
and
a
flat
policy
together.
How
do
you
understand
the
total
affecting
the
end?
A
How
do
you
even
go
about
designing
how
you
manage
and
group
your
different
resources,
so
we
can
effectively
apply
policies
to
them
so
now,
all
of
these
things
are
really
tricky
problems,
but
at
the
end
of
the
day,
why
do
we
even
care?
Is
it
okay,
if
it's
just
messy,
so
we
look
at
reality.
What
happens
when
it
is
messy?
A
Even
those
of
us
work
on
cloud
were
already
confused
by
this.
You
can
imagine
how
hard
it
is
for
customers,
so
as
an
enterprise,
admin
or
developer
tries
to
make
sense
of
these
things
first
day,
they
see
a
lot
of
inconsistencies.
They
say
they
need
to
do.
This
need
to
learn,
20
different
ways
of
doing
the
same
thing,
and
that
tends
to
be
you
know
it's
a
really
complex
system.
Nobody
fully
understands
it
when
it's
complex.
A
What
we
notice
is
that
two
types
of
things
happen:
either
people
do
nothing,
it's
it's
inertia,
they
just
decides
too
hard.
I'm
not
gonna,
make
any
change
right.
Oh
they
do
make
change,
but
because
they
don't
understand
there,
they
misconfigured
the
system.
What
typically
happens
is
when
a
developer
goes
to
an
arming,
set
and
says:
hey
I
need
access
to
the
storage
bucket
in
the
Army
says:
ok,
I've
made
the
policies
I've
made
a
policy
changes,
I
think
I
need
to
make
for
you.
Can
you
try
it?
A
They
go
back
and
says
it
doesn't
work
and
three
times
later
and
when
a
student
work,
the
a
being
out
of
goodness
of
their
heart,
want
to
unblock
a
developer.
Well,
you
should,
you
should
say:
okay,
I'm
gonna
make
the
bucket
public.
That's
how
we
end
up
seeing
so
many
security
incidents.
Do
you
know
sensitive
data,
storing
public
packets,
so
I
did
a
quick
search
and
then
you
know
they
are
lead
to
really
kind
of
you
know:
loss
of
control
and
and
breach
of
sensitive
data.
A
So
when
I,
when
I,
did
a
quick
search
on
internet
looking
for
articles
related
to
kind
of
inconsistent,
figurations
and
security
breaches,
there's
no
shortage
of
these
stories
right,
a
lot
of
cloud
users
already
feeling
like
it's
too
complex,
we
need
a
unified
stack.
We
just
count
to
manage
so
many
inconsistent
and
complex
configurations
and
policies,
and
then,
when
we
do
configure
them
around
there's
no
shortage
of
data
breaches.
It's
no
long.
It's
we
almost
getting
getting
to
the
point
of
being
numb.
A
There's
every
month,
there's
yet
another
headline
about
data
breaches,
so
my
ask:
is
we
collectively
need
to
figure
out?
What
are
we
going
to
do
about
it
because
eddying
of
the
day
when
we
get
to
such
a
complex
and
insecure
system,
we
are
suffer.
It's
not
a
about
a
single
company
of
single
cloud
or
iCloud
in
general,
gets
a
bad
reputation.
You'll
probably
already
have
you
know
you
either
people
your
own
company
or
your
customers,
saying
to
you
I'm,
not
really
sure.
B
Another
dimension
that
I've
found
really
challenging
is
sort
of
the
config
validation
type
policy
versus
the
more
like
runtime
enforcement
policies.
So
as
an
example,
something
like
admission,
control
in
kubernetes
versus
network
policy,
which
is
applied
to
a
running
container
and
I,
found
it
challenging
to
kind
of
reconcile
those
models
and
think
about
how
they
could
compose.
Well,
anybody.
A
Else
have
encountered:
config
validation,
bitanga
your
mail.
Oh
sure,
you
know
you
say:
policies
you're
sure
take
a
push
across
a
number
unconscious
machine,
I
assure
watching
a
neon
JC,
a
container
hotel
with
a
network.
Wouldn't
your
external
IP
make
sure
you
see
on
a
policy
that,
in
my
experience.
A
So,
to
extend
what
Tim
was
talking
about?
Well,
we,
what
we
find
is
there
are
policies
that
are
very,
very
much
about
whether
you
are
a
you
know
whether
you
are
able
to
access
particular
resource
or
to
call
particular
API
typical
I
am
our
back
title
policies.
That's
one
side
wouldn't
have
a
policy.
The
other
type
of
policy
has
become
very
popular
and
I.
Think
no
OPA
has
been
doing.
Some
work
in
that
area
is
about
configuration
validation,
so,
yes,
you
can
create
a
namespace,
but
only
with
certain
configurations.
Right,
yes,
you
can.
A
You
know
you
can
set
up
your
network
configurations,
but
only
within
these
four
three
choices
so
to
validate
certain
configurations
can
be
technically
correct,
but
for
your
particular
environment
for
your
security
requirements
for
your
origin
requirements,
they're
not
designed
by
the
administrators
I,
don't
have,
has
anybody
else
encounter?
These
kind
of
so
we've
noticed
is
in
some
in
some
it's
kind
of
a
spectrum
right.
There
depends
on
culture
in
some
companies,
it's
very
bad
enough.
That
developers
have
a
lot
of
freedom,
so
they're
told
to
just
go.
A
You
know
cradle
namespaces,
cradle
and
clusters,
create
your
own
services
around
whatever
you
want,
and
the
admins
will
observe
them
aside
or
publish
some
best
practice
guidelines,
hoping
the
developers
were
to
hear
to
them
in
the
other
extreme,
we're
also
seeing
environments
where
it's
very
it's
very
top-down.
The
armies
go
to
control
everything
when
developers
need
to
go
and
deploy
something,
especially
when
it
seemed
that
the
production
environment,
what
they
actually
need
to
do
is
to
create
a
new
file
ticket
and
it
goes
through
a
system.
A
Because
from
from
my
point
of
view
of
seeing
that
more
kind
of
in
Europe
and
the
US
and
I'm
sure
it's
starting
to
happen
major
as
well,
where
they
they
build
a
CI
CD
pipeline,
any
policy
and
config
deployment
as
well,
so
all
the
policies
no
sitting
and
central
repo
right
and
I'm
from
there,
and
whenever
you
need
to
make
any
change
to
the
policies
you
know
nobody
can
touch
production.
Only
a
service
account
can
automatically
deploy
things
to
production.
A
So
then,
what
you
do
is
to
to
push
any
changes
for
you
to
get
repo
and
then
from
there
on
goes
just
goes
through
the
same.
You
know:
request
approval,
workflow,
the
CI
CD
pipeline
to
eventually
go
to
production.
Now
that
kind
of
framework
creates
a
space
for
doing
you
know
configured
configuration
validation,
kind
of
policies
right
so,
as
you
can
imagine
anybody
who's
familiar
with
CI
C
d-four
code
and
binaries
they're
things
like
resubmit
checks
right,
you
can
check
whether
the
code
is
even
valid.
Is
you
know
static
analysis?
A
You
can
do
that
at
a
point
when
the
code
gets
into
repo
or
when
a
code
gets
broke.
You
know
your
initial
dev
branch
into
production
branch
and
then,
when
we
do
that
for
policies
that
also
gives
us
an
opening
a
place
where
we
can
do
some
similar
status
statistic:
analysis
policy,
for
example,
you
can
you
know
some
companies
can't
even
talk
about
things
like
we
want
to
name
names
faces
the
same
way
right,
there's
a
certain
format.
We
name
them.
A
C
Thank
you,
yeah
I
mean
you've
touched
on
basically
the
kind
of
things
we
do
as
well,
just
to
I
put
some
more
flesh
and
those
things.
Certainly
the
difference
between
this
will
network
policy,
which
can
only
really
be
validated
by
dynamic
test
against
cluster
like
statically
analyzing.
Those
routes
is
really
difficult
versus
things
like
pod
security
policy
and
I
guess
our
back
or
that
kind
of
falls
into
the
dynamic
test
as
well,
because
you
can
kind
of
that.
Our
guess
is
a
gray
area
there.
But
again
we
dynamically
test.
C
Those
but
in
terms
of
the
of
commit
to
a
git
repository,
have
a
pull
request
to
have
that
checked
over
by
somebody
with
centralized
access.
We
then
run
static
analysis
on
Yammer,
which
just
kind
of
feels
weird,
but
ultimately,
with
this
fully
declarative
system
like
where
else
do
you
put
those
sanity
checks?
So,
for
example,
like
we
actually
make
sure
that
if
someone
wants
to
make
a
change
to
a
pull
request
to
the
configuration
repo
for
a
given
namespace
and
they
change
the
port
security
port?
C
So
there's
the
system
configuration
for
that
as
opposed
to
the
developer
configuration
and
the
division
there
is
the
system.
Stuff
is
what
the
SRE
team
control
I
mean
in
our
model,
for
this
versus
the
developers
have
access
to
change
things
like
network
policy,
for
example,
but
if
they
want
to
make
a
change,
the
pod
security
policy
that
has
to
go
back
in
through
the
system
repo
and
then
see
how
actually
checks
that
they
haven't
turned
off
hosts
the
disabling,
the
hosts
sharing
of
namespaces
and
privilege,
and
things
like
that.
A
And
that's
a
great
point
like
the
when
you're
making
changes
to
policies
a
lot
of
time.
It's
not
about
a
single
persona
right,
you
might
have
multiple
admin
teams.
When
one
managers,
the
actual
attack
policy's
versus
some
other
started,
networking
policies
you
guys
have
developers
will
may
be
a
bit
of
control
policies
to
talk
suriname
extent
for
their
own
namespaces.
So
when
you
have
multiple
people
collaborating
that's
also
where
you
know
the
git
workflow
actually
is
built
for
people
to
collaborate.
A
You
can
do
request
approval
workflows,
you
can,
you
know,
kind
of
fork,
branches
and
merge
back
and
and
we
can
really
kind
of
leverage
the
same
existing
workflows
for
the
policy
world
as
well.
So
I'm
just
gonna
throw
out
a
few
a
few
ideas.
So
you
know
when
I
look
at
this
long
list
of
policies
trying
to
think
how
we
can
make
make
sense
out
of
them.
Yeah
question:
oh
yeah
that'd
be
great.
Can
we
the
who
has
microphone.
D
Thank
you,
so
I
just
want
to
bring
up.
Another
idea
about
here
so
I
I
understand
is
a
inconsistency
between
these
security
policies
because
the
operates
and
the
different
layer,
some
of
them,
therefore
the
IPA
or
something
they
have
for
the
service
level
like
it
used
to
attack
the
cooperate
state
overseas
for
the
IP
level,
the
operating
systems
they
help
their.
That
is
right,
permission
or
refurbishing
for
files
label.
So
I
understand
this
in
inconsistency
because
they
have
given
abstract
here,
but
the
Christian
is.
D
Does
the
complexity
comes
from
the
inconsistency
because
essentially
they
are
different.
So
it's
it's
it's
it's
just
what
it
is.
So
it's
hard
to
find
a
unique
model
to
you.
Niagra
represent
different
security
policies
and
a
different
label.
So
could
it
be
possible
that
the
complicity
comes
from
the
peoples?
Are
operators
the
programmer?
We
do
not
have
a
full
understanding
from
the
lowest
level
to
the
absolute
wall.
We
are
good
and
like
like
one
specifically
for
someone
who
is
working
on
the
data
properties.
D
He
may
do
not
have
enough
knowledge
about
the
issue
up
and
policy.
So
could
it
be
possible?
We
can
mitigate
this
by
increase
absorbability
of
these
policies
like
we
have
a
standard
about
when
a
request
is
denied,
we
have
a
standard
response
for
now,
I
think
different
levels.
They
have
different
behavior
in
IP
level,
when
a
request
is
derided,
it
just
drops
a
packet
and
from
the
users
clarity.
A
So
one
of
the
you
know
when
we
saw
so
many
policies
we're
not
for
first
thought
it
was.
Is
it
just
because
we're
now
we're
not
being
responsible
in
creating
policies?
They're
too
many?
Can
we
consolidate?
Can
we
actually
make
them
the
same
policy?
Who
came
to
that
realization,
I
they're,
actually
different
layers?
That's
a
different
use
cases.
Wow
I
think
we
will
professionally
living
a
world
of
many
policies
serving
different
purposes.
So
I
you
know,
and
then
that's
where
I
think
some
of
the
ideas
we
start
developing
are
more
like
the
product.
A
There
will
be
many
policies.
There
will
be
different
policies.
The
things
we
can
make
consistent,
like
you
said,
rise
to
interesting
things
like
our
reporting
and
also
in
things
like
workflows
right.
If
we're
ever
gonna,
do
things
like
simulation?
Is
there
a
standard
interface
for
different
type
of
policies
to
plug
in
so
you
can
all
result
in
the
same
simulation
experience
for
for
the
end
it
happening
or
the
end
user.
A
D
A
Yeah,
so
then
we
can
talk
about
touching.
You
know
that
that
brings
us
to
another
thing
we
can
talk
about.
We
look
at
those
different
layers,
so
we
actually
thought
about.
Are
there
ways
to
make
the
syntax
a
little
simpler,
right,
annoying
they're,
different
policies?
Are
there
ways
jellies
to
bring
the
syntax
to
be
somewhat
singular,
to
have
a
framework
to
think
about
how
they
fit
together?
And
there
are
ways
to
standardize
workflows
to
standardize?
A
You
know
all
the
other
analytics
interface
as
well
on
this
on
a
syntax
layer
at
a
very
basic
level,
I
started.
Thinking
about
you
know,
CRTs
and
krm
is
a
great
interface
right,
so
sto
kubernetes,
where
everything
is
converging
into
you
know,
describing
your
resources
or
CRTs,
and
maybe
that's
another
thing
that
our
each
other
cloud
you
know,
other
different
clouds
are.
A
The
different
platforms
can
also
starting
converging
them,
because
if
we
can
converge
on
a
similar
way
to
describe
resources,
then,
on
top
of
that,
you
know,
the
existing
policies
like
Arabic
and
OPA
can
potentially
work
across
multiple
platforms
and
we've
thought
about
like
our
back
is
a
great
type
of
policy
for
kind
of
more
more
coarse-grained
aadmi
operations.
Right,
if
you
you
know,
if
you
fit
into
one
of
the
standard
verbs,
it
describes
permissions
for
you
pretty
well
and
then,
when
getting
to
data
access
that
that
tends
to
get
more
granular.
A
So
people
want
to
do
that.
That's
not
only
by
our
buy
policies,
but
also
bad,
for
example,
could
be
in
contact
to.
Where
are
you
coming
from
the
light
IP
address,
I'd
like
em
of
the
day
or
depending
on
the
content,
the
content
of
the
data?
So
you
can
say
you
know
you
have
access
to
the
table,
but
only
if
the
column
made
me
call
the
flower
or
the
value
could
be
black
right,
so
you
can
filter
out
the
most
sensitive
data,
so
we're
seeing
you
know
the
trend.
A
And
then
we
can
talk
about
what
happens
in
higher
layering
the
management
layer
right,
so
the
gay
Doc's
experience
does
seem
to
be
we've
seen
a
lot
of
customers
who
native
they
start
to
move
towards
a
kidnaps
experience
for
managing
the
many
different
policies.
So
you
can
have
different
types
of
policies.
You
can
even
have
policies
for
different
platforms,
however,
get
the
source
of
truth
is
where
you
can
actually
have
a
lot
of
them
store
together
and
essentially
manage
together,
and
then
there
you
can
also
separate
by
different
branches.
A
Nobody
has
done
great
job,
analyzing
and
explaining
policies
of
stimulating
policies,
so
that
I
think
is
the
next
very
challenging
problem
that
we
have
to
you
know.
So,
together
are
there
so
I?
My
knowledge
might
not
be
complete.
Are
there
other
existing
kind
of
components
in
ciencia
that
are
to
do
with
how
we
analyze
and
simulate
and
explain
policies?
Does
anybody
know.
A
So
that
might
be
an
opportunity
you
know
for
us
to
to
to
think
about
what
we
can
be
able
to.
You
know
make
managing
these
management
of
these
policies
in
color.
I.
Think
standardizing
error
message
is
a
very
simple
thing
and
prairie
low-hanging
fruit,
I
can
get
a
lot
of
results
already,
would
love
to
see
if
they
are
satsang.
You
know
how
we
convert.
Yes,
ideas
are
back.
E
C
A
A
We
were
exciting
too,
if
we,
you
know,
start
working
on
something
like
that
in
CNCs.
I.
Think,
though,
there's
a
lot
of
desire
for
that
when
we
talk
to
customers,
we
use
you
know,
say:
I
am
on
GCP
a
lot
of
times.
We
ask
them.
What
do
you
like
a
simulator,
there's
a?
Why
haven't
you
build
one
last
year
already
I'll
totally
pay
for
it.
It's
you
know
would
be
one
of
the
most
useful
tools
and
then
another
idea
on
the
other
side.
C
The
answer
to
that
is
maybe
sort
of
like
centralized
log
with
a
specific
are
back
permission
to
read
just
and
I
don't
know,
and
the
other
thing
is
like
on
the
nature
of
simulation,
the
the
auerbach
can
I
like
interface
for
cube
control
is
it's
like
super
powerful
and,
like
obviously,
the
the
basis
for
testing
those
things
dynamically,
so
it
may
be.
Some
sort
of
kind
of
cumulative
can
I,
which
then
hits
all
the
policy
points
and
tries
to
bring
back.
Something
about
could
be
that
are
going
some
way
towards
that
simulation.
C
A
A
However,
things
from
you
know,
results
and
errors
from
your
policy
checks
actually
often
carry
sensitive
information
right
there,
either
PII
information
sensitive
information
to
your
applications,
or
they
can
also
be
things
that
a
hacker
can
use
to
to
try
to
bypass
the
sister
and
figure
out
how
to
hacking
to
assistance.
So
there
is
a
pool
between
you
know
the
more
information
the
better
for
analytics,
but
it's
it
also
to
time
exchange
and
weakening
your
system.
A
Your
defense,
as
well
and
I,
think
that
you
know
brought
up
a
very
good
point
like
if
there
are
places
where
we
can
aggregate
those
information
by
the
the
wrong
information,
something
you
don't
want
and
then
a
user,
a
hacker
actually
to
be
able
to
easily
get
a
hands
on
Harvard.
Did
the
aggregating
information,
the
the
final
recommendation
or
suggestions?
A
That's
where
we
can.
We
can
kind
of
hide
the
sensitive
information
right,
make
it
more
about
the
policy
you're
looking
at
rather
than
you
know,
the
data
that
might
get
leaching
in
the
errors
are
logging
is
another
interesting
thing
here.
We've
seen
kind
of
you
know.
First,
you
have
outta
logs
and
people
say
I,
don't
know
enough,
and
then
you
start
developing
more
granular
other
logging
policies
as
well
right.
A
You
want
to
say,
for
example,
only
outer
log,
the
user
accesses
to
to
these
resources
and
not
a
service
account
accesses,
because
those
are
tend
to
be
noise.
Those
tend
to
be
just
system
automation
and
then
that
becomes
in
another
place.
Somebody
can,
potentially,
you
know,
try
to
access
something
but
turn
out
the
other
logging
and
give
them
the
way
out
of
the
system
as
well.
So
yeah,
that's
a
fascinating
problem.
So
so,
when
we,
when
we
thought
about
you,
know
simulation,
how
do
you
explain
a
policy?
A
number
of
models
seem
possible
to
explore.
A
I
think
you
know.
You're
gonna
have
an
emulator
Ryan
a
lot
of
different
policies
to
plugging
and
give
you
the
end
result
of
the
simulation
is
one
option.
Well,
Russell
thought
about
you
know:
sin
boxing
just
create
easily
recreate
environment
and
then
be
able
to
run
your
policy
through
that's
another
option.
Another
option
that
I've
come
up
is
to
rerun
your
production
traffic.
A
So
if
you
can
capture
a
production
traffic
from
the
last
hours
days
right
and
then
you
can
identify
representative
traffic
and
then
round
them
through
the
new
policies,
and
then
you
can
tell
without
turning
the
policies
down,
you
can
simulate
into
how
what
without
you
might
get,
and
then
we've
also.
You
know,
of
course,
with
machine
learning
anytime.
The
problem
gets
contacts.
A
We
always
wonder
what,
if
we
throw
a
little
machine
learning
on
it,
is
it
possible
to
identify
again
the
typical
traffic
and
a
typical
responses
to
the
policies,
so
yeah
I
mean
I
would
love
to
see.
You
know,
that's
something
I
think
every
everybody
I've
talked
to
a
lot
of
different
companies.
Are
the
different
kubernetes
uses.
I've
talked
to
that
everybody's
very
early
on
in
analyzing,
simulating
policies
and
that's
an
area
would
love
to
see
some
development
in
CNC
F.