►
From YouTube: Kubernetes Federation WG sync 20180131
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
B
A
I,
don't
know
about
like
the
trade-offs
for
like
annotations
versus,
like
top
level
like
specific
fields
versus
like
a
separate
resource.
I.
Think
that
that's
like
from
my
perspective,
the
policy
engine
should
be
able
to
work
with
any
of
those
hoods
and
I.
Think
OPA
can
today
so
I'm,
not
I.
Don't
have
too
many
opinions
about
that,
but
yeah
about
like
where,
like
I
just
VI
on
my
main
question,
was
just
around
like
what
the
goal
was
of
that.
A
E
C
B
B
What
we
had
sort
of
as
an
annotation
face
mechanism
into
this
proposal
and
I
think
Turin
had
some
concerns
and
it
kind
of
made
me
wonder
whether
I
really
understood
how
policy
wasn't
like
what
his
purpose
was.
What
you
know
there
may
be
existing
implementation,
represented
like
an
ideal
like
we're
just
gonna
annotate
things
and
provide
a
per
resource
indication
of
what
should
happen.
Or
should
there
be
something
else
so.
C
Now
I
mean
I
turn
can
talk
to
in
more
detail.
My
perception
is,
we
just
kind
of
had
something,
quick
and
reasonably
dirty
together
to
illustrate
the
principle
and
wasn't
really
designed
in
any
detailed
way.
I
mean
I.
Think
turn
you'd
knocked
that
together
in
a
couple
of
from
what
I
recall
so
yeah
I
think
we
can
certainly
step
back
and
say
how
should
policy
fit
in
here
better
how
much
work.
F
A
B
Like
the
you
know,
the
selector
mechanism,
I've
kind
of
just
assumed
saucers,
are
kind
of
a
cube
native
thing
that
people
would
understand.
I
understand
that
they're,
limited
and
I
guess
I
was
expecting
that
someone
would
be
able
to
define
placement
fee,
like
selectors
I
sure,
like
a
like
a
low-hanging
fruit
type
of
solution,
and
if
they
needed
something
more
complicated,
they
would
set
the
policy,
but
I
was
naively
assuming
that
policy
would
sort
of
just
translate
into
something
like
selectors
and
from
what
from
their
comments.
Maybe
that's
not
the
case.
A
C
I
think
what
I
would
personally
like
to
get
out
of
this
meeting
is
some
concrete
stuff
that
that
people
can
go
off
and
start
milling.
Pr's
and
I've
got
a
few
kind
of
ideas,
but
maybe
if
one
has
others
as
well
and
Maru
so
I'd
like
just
like
to
make
sure
we
get
there.
At
least.
Maybe
we
can
time
box
a
policy
discussion
to
make
sure
that
doesn't
like
backburner
for
too
long
happy
to
chat
for
chat
about
it
for
ten
minutes
or
five.
Let's.
C
A
Placement
policy
stuff
and
then
we
can
move
on
other
topics.
Okay,
so
I
guess.
Yeah
summary,
you
asked
one
question
which
was
like:
what's
the
relationship
of
like
our
back
policy
or
like
authorization
policy,
and
it's
been
policy
and
how
OPA
gets
into
that
picture,
and
so
just
to
like
address
that,
because
that's
kind
of
a
high
level
thing
like
oppa's,
basically
designed
to
allow
you
in
different
kinds
of
policies
throughout
your
system
at
different
layers
of
the
stack
and
so
what
it
gives.
A
You
is
this
policy
language
that
you
can
use
to
encode
like
rules
about
how
the
system
should
behave.
It
isn't
coupled
any
particular
like
project
or
layer,
the
stack
or
domain
or
anything
like
that.
Okay,
so
you
can
use
it
to
enforce
authorization
policies.
If
you
want,
you
can
use
it
to
enforce
placement
policies,
you
can
use
it
for
admission
control
policies,
there's
all
kind
of
different
applications
for
it.
So
it's
meant
to
be
general.
A
B
Think
so,
I
the
reason
I
was
bringing
a
bar
back
is
having
placement
be
on
a
resource
versus
like
it's.
A
separate
resource
would
allow
like
NaCl
to
be
enforced
via
our
back
sure,
but
if
we're
talking
about
like
an
admission
controller,
that's
using
OPA
or
like
something
that's
more
specific
than
maybe
you
don't
want
to
use
our
back
and
you
could
apply
like
you
can't
change.
This
feel
what
I
was
concerned
about
was
if
there
was
an
administrative
policy
that
said
replica
sets
with
these
characters
they
Sneed
to
go
here.
B
You
know
to
these
Busters
if
it's
selector,
based
or
if
it's
just
filled
base
and
there's
no
protection
on
the
fields
that
are
being
set
and
someone
can
go
in
and
override
the
policy.
So
there
has
to
be
a
way
of
determining
what
the
policy
should
be
and
then
making
sure
that
it's
enforced,
I
guess
sure.
A
Yeah
and
and
I
think
that,
like
so
I
think
that
access
control
over
the
policies
themselves
is
interesting,
but
it's
it's
sort
of
like
I.
Imagine
stout
us
out
of
scope.
Two
stomachs
like
that
today
in
Federation
I,
don't
know
what
it's
called
map
federation,
1.0
or
whatever
that
in
that
model
you
you
load
the
the
placement
policies
as
config
maps
and
you
could
express
you
can
enforce
access
controls
over
those
config
maps.
Today,
if
you
wanted.
B
A
You
asking
like
whether
specific
developers
should
be
allowed
to
request
certain,
like
just
specify
certain
certain
pieces
of
intent
at
all,
is
that
is
that
what
you're
asking
like
like
today
in
in
in
1.0,
the
user,
would
like
deploy
an
application,
and
they
would
say,
like
this.
Application
requires
PCI
compliant
clusters,
or
it
would
require.
You
know
like
a
u
jurisdiction
or
something
like
that.
Right,
alright,
and,
and
so
that
would
be
sissified
by
annotations
or
labels
or
whatever
or
other
fields
in
the
mat.
Just.
B
That,
if
okay
so
you're
in
your
specific
example,
if
I
had
a
replica
set
that
said,
I
require
PC
PCI
compliant
clusters
like
something
way
of
indicating
that
and
the
policy
interpreted
that
and
said.
Well,
then
you
should
be
placed
in
these
clusters.
I
would
put
that
placement
information
like
in
the
current
iteration.
B
F
Whoever
gets
the
access
to
that
particular
object.
In
my
opinion,
you
should
be
the
case
right.
Okay,
so,
for
example,
user
has
created
replicas
set
in
a
particular
namespace
and
he
has
access
to
only
in
space
and
he
can
access
that
replica
said
then
that
user
would
be
able
to
update
or
change
it
also.
C
What
policies
intended
to
to
not
allow
so
so
policy
might
say
you
know
you,
your
user
X
and
you
don't
get
to
actually
choose
which
clusters
you
go
in.
We
choose
for
you,
you
can,
you
know,
give
us
some
basic
requirements
with
PCI
or
you
need
this
for
that.
But
but
you
can't
actually
choose
where
to
put
your
stuff,
because
we
decide
which
postures
you
going
to
based
on
costing
and
whatever
policy
might
apply
so
I
think
that's.
C
The
key
thing
is
that
there
that
the
person
creating
the
resources
is
kind
of
intentionally
different
than
the
person
creating
the
policy
and
the
person
creating
the
resources
fundamentally
has
different
permissions
than
the
thing
enforcing
the
policy
and
and
I.
Think.
If
I
can
paraphrase
maru's
question,
how
do
we
actually
like
implement
that?
How
do
we
make
sure
that
that
the
policy
engine
can
execute
policy
decisions
and
get
through
the
necessary
our
back
rules
that
may
be
in
place
to
execute
that
but
model?
C
How
the
created
the
resource
to
be
able
to
do
the
same
thing
and
I
think
we
can
I
mean
I
would
suggest
we
sort
of
climb
boxes
somewhere.
I
think
we
should
we,
you
need
to
solve
that
problem.
I,
don't
think
it's
like
a
fundamental
problem.
I
think
it
may
so
there
are
a
couple
of
different
ways
of
thinking
about
it.
None
of
them
I,
think
material
massively
affected,
but
design.
Overall,
the
one
is
field-based,
our
back.
Where
and
I
know
we
don't
have
that
yet
and
I
know
we
want
that
etc.
C
Don't
know
what
a
general
term
we've
mentioned
for,
even
if
you've,
given
a
general
term
that
we've
cut,
we
would
overrides,
we
got
placement
and
we've
got
propagation.
We
got
all
of
these
things
which
are
looking
like
they're,
a
sort
of
constellation
of
resources
around
a
primary
resource
that
needs
to
go
in
through
one
or
more
clusters
and
by
appropriately
factoring
those
we
can
say.
C
Well,
you
know
some
of
them
are
accessible
to
the
policy
enforcement
engines
which
made,
for
example,
the
term
in
which
clusters
it
goes
into
may
even
propagate
into
the
clusters,
and
some
of
them
are
accessible
to
the
user.
Who
wants
the
resource?
You
know
iterated
and-
and
you
may
even
have
like
some
kind
of
an
multiple
inheritance
sort
of
override
thing
where
you
know
what
the
policy
design
overrides,
what
the
user
asked
for
I.
B
Mean
I
think
we
can
probably
take
this
offline.
It
sounds
like
minutes
to
being
a
more
involved
discussion.
The
reason
I
was
kind
of
concerned.
We
discussed
this
as
soon
as
possible.
Is
that
I
think
there
is
kind
of
a
problem
with
the
existing
like
with
the
proposed
way
of
using
solicitors,
because
if
you
want
to
protect
like
what
a
policy
applies
to
say,
I
want
to
have
these.
You
know
I
want
to
say
these
reserves
go
to
these
clusters.
A
B
A
I,
remember
that
that's
a
really
that's
a
good
point.
I,
think
that,
like
I
think
that
in
a
high
level
like
I
could',
was
saying
earlier
that,
like
you
know,
we
did
this
work
really
quickly
in
100,
but
I
think
that
at
a
high
level,
the
decision
that
was
made
to
decouple
like
the
placement
decision
logic
from
the
actual
like
enforcement,
was
like
a
really
good
one,
and
I
think
that
that's
a
good
model
going
forward
because
it
allows
anybody
to
plug
in
their
own
their
own
placement
logic
that
they
want.
A
A
You
yes-
and
that
was
my
main-
that
was
my
main
thing
and
I
wanted
to
understand
within
the
proposal,
because
in
the
proposal
it
looks
like
that
decoupling
was
going
away
and
we
were
going
to
be
left
with
this
federated
placement
object
that
contains
a
miniature
DSL
for
describing
how
workloads
should
be
placed
onto
clusters,
and
that
was
the
that
was
the
kind
of
core
point
that
I
was
trying
to
understand
so
happy
to
have
a
follow-up
conversation
to
go
into
this
in
more
detail.
Yeah.
B
C
As
a
brief
aside
before
we
leave
this
topic
completely,
I
am
trying
to
remember
exactly
what
context
this
came.
I
think
it
was
in
the
context
of
multi-tenancy.
Actually,
there
was
similar
problems
related
to
you
know,
interaction
with
resources
by
multiple
tenants
and
how
limited
it
was
and
where
the
namespaces
for
the
right,
abstraction
and
house,
isolated,
namespaces
were
and
all
that
kind
of
thing,
and
it
sort
of
has
a
lot
of
commonality.
C
With
the
problem
we're
facing
here
and
one
thought:
I
had
there
and
I
haven't
turned
it
into
a
full
design
proposal,
but
I
think
it's
got
quite.
A
lot
of
potential
is
to
is
to
actually
elevate
labels
to
like
first-class
objects,
so
labels
are
owned
by
somebody
and
you
can
apply
our
back
rules
to
them,
and
you
can
say
this
person
can
read
that
label
and
person
can
write
that
label
and
this
person
can
attach
that
label
to
to
this
thing,
and
then
many
of
these
problems
become
much
more
tractable
I
think
so.
C
C
No,
the
proposal,
the
proposal
was
that
it
would
be
a
general
mechanism
for
solving
this
class
of
problem
and
and
and
it
was
basically
a
couple
of
so
one
thing
was
like
first
class
labels,
which
you
could
apply
our
back
to
and
the
other
one
was
having
them
proper
having
them
hierarchical.
So
you
could,
you
know,
apply.
C
Rules
are
back
rules,
two
kind
of
hierarchies
of
labels,
so
you
can
say
anything,
that's
a
sub-label
of
this
thing.
You
know
it's
not
readable
by
these
people
and
it
turns
out
to
be
a
reasonably
powerful
model,
but
but
I
need
to
write
it
down
in
sort
of
more
detail.
I'm
sure
it
can
be
ripped
to
shreds
by
other
people
as
soon
as
I
write
it
down,
but
it
seems
worth
writing
down
anyway.
Even
if
we
get
the
objections
to
it,
formalized
in
a
set
of
comments.
B
A
B
Like
that,
they
can't
just
be
wiggled
out
of
by
changing
tables
by
anybody.
Maybe
that's
at
least
one
problem
to
solve,
and
so,
if
you're
as
long
as
you're
clear
in
the
problem
statement,
I
think
putting
that
down
on
paper
gives
people
an
opportunity
to
explore
that
solution,
but
also
consider
other
solutionist
yeah.
D
B
A
E
B
C
F
C
F
Okay,
let's
move
ahead,
then
the
last
meeting
when
we
were
off
the
last
meeting
I
think
we
were
at
the
discussion
around
placement.
So
one
thing
was
concluded
that
we
keep
for
placement
like
Turin
was
also
saying
right
now
that
50
placement
logic
decoupled
from
the
actual
object,
and
so
currently,
whatever
is
proposal.
The
document
it
seems
to
be
right.
F
C
Yeah
I
think
I
think
we're
heading
in
the
right
direction.
One
brief
kind
of
thought,
I
had
and
I
won't
take
more
than
30
seconds
here.
I
would
like
us
to
just
bear
in
mind
like
what
the
baseline
is
for,
like
ease
of
use,
which
I
think
is
the
current
system.
You
know
you
can
just
take
a
normal,
familiar
kubernetes
object,
you
can
create
it
and
it
works,
and
you
can
then
you
know
it
works
just
on
extent,
and
then
you
can
incrementally
buy
at
the
moment
applying
annotations.
C
You
can
influence
where
it
gets
placed
and
some
other
things
now.
It
has
limitations
and
we
nobody's
claiming
it
it's
great
but
but
I
think
that's
kind
of
like
a
baseline
for
ease
of
use
and
I
think
we
should
try
and
make
sure
that
we
always
score
whatever
other
proposals.
We
have
against
that
baseline
just
to
make
sure
we
don't
diverge
too
far
away
and
end
up
with
something
which
is
like
massive
MIMO
other
people
to
use
and
then
the
baseline
and
then
now
the
call.
C
I
guess
I
guess
the
sort
of
thing
that
made
it
crossed
my
mind
right
now:
we've
got
propagation,
we've
got
class.
Forget
all
the
terms.
We've
got
status,
we
got
propagation,
we've
got
placement,
we
don't
have
we
sort
of
have
some
sort
of
template
thing,
although
that's
more
placements
and
waiting
and
all
of
that
stuff.
So
we
have.
C
You
know
a
fairly
substantial
subset
of
what's
spoken
about
here
implemented
and
it's
relatively
straightforward
to
use
and
in
the
simplest
case
you
can
actually
just
create
a
communities
resource
and
does
waiting
and
default
propagation
and
etcetera
totally
and
and
in
order
to
get
that
sort
of
behavior
with
a
new
system.
You
end
up
having
to
create
like
seven
different
things
and
apply
labels
to
tie
them
all
together
and
make
sure
that
they
all
hooked
up
together
properly.
B
Of
you
know
compose
basically
things
out
of
different
pieces,
but
I
think
it
likely
that
anything
that
is
tied
directly
to
a
resource
like
overrides
I've,
been
kind
of
like
implementing
like
overrides
and
I'm
like
well.
This
is
kind
of
annoying
like
the
implement
is
a
separate
resource
and
then
I
have
to
have
a
relationship
between
the
two
things
and
I
mean.
A
B
There's
a
policy
to
like
remove
one
if
one's
deleted
or
not
or
it
just
yeah
so
I
mean
to
me
like
overrides,
is
probably
the
biggest
candidate
for
merging
and
having
not
just
be
part
of
the
template
resource
placement.
I.
Think
that
there's
advantages
to
having
placement
separate
that
we
discussed
last
meeting
so
at
least.
C
C
I,
don't
have
any
objections
to
the
advantages
of
the
disconnecting
these
things.
I
just
want
to
make
sure
that
we
don't
get
carried
away
with
factorization
and
then
very
difficult
to
use
systems
and
then
I,
guess
related
to
that
concern
is
we
can
always
argue,
then,
oh,
but
you
can
just
create
a
layer
above
this
thing
and
hide
all
of
these.
You
know
complex
relationships
and
separate
objects
by
creating
a
sort
of
a
different
higher
level
abstraction
which
ties
them
all
together.
C
But
if
that
is
the
case
there's
and
in
order
to
use
that
higher
level
of
abstraction,
you
have
to
hand
all
of
the
underlying
stuff
anyway,
and
it
has
to
be
you
know,
specific
versions
and
specific
controllers
and
things
that
do
all
the
right
stuff,
in
which
case
like.
Why
did
you
make
them
all
so
pluggable?
C
In
the
first
place,
if
all
you
did
was
fill
the
layer
on
top
of
them
and
you
need
all
of
the
plugins
to
be
precisely
as
defined
by
the
higher
level
there,
maybe
you've,
just
kind
of
created
a
whole
lot
of
anything.
That
was
just
a
thought
as
long
as
we
keep
that
in
mind
and
don't
end
up
accidentally
at
a
place
where
people
have
to
create
five
resources
for
the
registry
create
one
to
do
something
reasonably
straightforward.
B
So
I
should
say
that
I'm
pretty
allergic
to
complexity,
which
means
I,
might
have
a
mild
allergic
reaction
whenever
I'm
touching
cute,
but
okay
in
terms
of
the
implementation,
it's
like
I
think
the
only
way
that
you
can
create
what
you're
talking
about
is,
if
you
didn't
have
to
implement
it,
and
anybody
who
actually
has
to
implement
like
a
bunch
of
different
resources
and
deal
with
all
the
complexity
of
like,
like
it's
not
simple,
to
create
api's
in
queue
even
with
generators.
So,
to
my
mind,
it's
like
going
through
this
process.
B
B
B
A
good
thing
because
it
prevents
the
sort
of
small
that
you're
talking
about
I,
definitely
am
concerned
about
the
user.
Experience
too
I
mean
I
want
this
to
be
simple,
I
think
that's.
As
you
say,
the
existing
solution
sets
a
bar
and
if
we
go
too
far
beyond
that,
we're
gonna
be
harming
ourselves,
but.
C
C
So
I
would
argue
that
we
really
like
can't
afford
to
go
any
worse
than
that,
because
things
are
just
gonna
get
way
worse,
but
in
it
and
and
the
same
goes
for
configuration,
I
mean
or
rather
deployment
right
now
we
have,
you,
know
the
controller
manager
and
the
API
server
and
you
deploy
them
into
a
cluster,
and
people
can't
even
get
that
right
if
we
now
have
them
having
to
deploy.
You
know
a
dozen
of
these
different
controllers
to
make
stuff
work
makes
me
nervous
anyway,
I
mean.
B
B
Like
aggregation
is
gonna
be
painful,
at
least
initially
like
will
automate
it
make
it
friendly
for
the
user,
but
we're
moving
from
deploy
a
piece
I.
A
pea
I
can
API
server
and
controller
manager,
we're
moving
from
that
to
you
know,
deploy
and
aggregate
cluster
registry,
potentially
if
it's
not
redeployed
and
deploy
an
aggregate
like
the
API
server
and
play
the
controller
manager.
So
we're
already
looking
at
like
a
more
complicated
world
and
the
goal
is
just
going
to
be
to
automate
that
and
hide
it
from
the
user
as
much
as
possible.
B
F
Guys
I
was
I
mean
I
did
not
interrupt
you
in
between,
but
I
think
we
have
been
talking
about
the
functional
areas.
These
five
things
that
we
are
talking
about,
they
have
been
I
mean
I,
was
corrected
upon
the
and
I
assume
that
these
five
are
going
to
be
practical
resources.
So
we
haven't
concluded
that
we
are
going
to
have
five
different
resources
right
so
so
far
we
have
only
concluded
that
we
will
have
a
federated
object.
These
are
this
Copernicus
object
and
probably
a
placement
object
separate
from
it.
A
Yeah
function
of
the
of
the
input
so
like
did
you
have
a
if
you
have
a
like
a
replica
set
coming
in
and,
and
you
want
to
make
a
placement
decision
based
on
some,
you
know
old
and
not
any
like
some
arbitrary
field
in
that
replica
set
or
based
on
external
context,
for
example,
that
may
be
not
in
the
replica
said:
it's
changing
externally
ik.
How
would
you
how
do
you
achieve
that
with
this
proposal?.
B
Are
you
talking
about
like
policy
engine
deciding
placement?
Are
you
talking
about
like
the
dynamic
scheduling
because
I
think
like
if
there's
a
resource
that
needs
to
be
created
like
versus
a
field?
You
can
just
sum
up,
and
you
just
be
responsible
for
creating
a
policy
engine
would
be
responsible
for
creating
the
resource
and
setting
appropriate
field,
but
already
its
apartment.
A
B
Guess
the
question
or
which
is
what
you're
saying
raises
the
question
for
me:
if
the
policy
engine
needs
to
be
able
to
make
a
decision
about
placement
and
provide
a
placement
directive
to
the
propagator,
where
does
it
do
it
and
in
the
existing
model?
It's
just
the
object
that
it's
deciding
policy
for
has
a
field.
It's
just
set
yeah.
If
we
have
a
placement
resource
that
is
somehow
using
a
selector.
B
B
A
B
Overrides
are
separate,
though,
and
so
is
scheduling.
They
have
to
be
tied
directly
to
the
resource.
At
least
overrides
I.
Think
I
think
it
makes
sense
to
I
mean
you
could
have
it
differently,
but
I
mean
overrides,
have
to
be
typed
and
they
have
to
have
a
reference
to
what
they're
applying
to,
and
so
it's
like
just
extracting
the
fields
and
allowing
some
variation
but
places,
but.
C
Just
just
I
hope
I'm
using
the
correct
terminology
here,
so
so
a
user
specifies
they
want
a
bunch
of
stuff,
and
that
includes
like
placement.
Directives
like
I
would
like
to
be
placed
in
Europe
or
Asia
or
whatever,
and
they
specify
what
the
template
of
their
thing
looks
like,
and
they
specify
that
you
know
certain
overrides
should
take
place.
C
For
example,
when
you
put
stuff
in
Asia,
you
should
substitute,
you
know,
give
it
the
Asian
database
or
whatever
it
happens,
to
be,
but
then
there's
actual
controllers
doing
this
stuff.
So
there's
a
there's
some
logic
somewhere
that
that's
choosing
with
or
without
the
input
of
policy
to
decide
like
which
asian
cluster
to
choose
and
and
and
that's
got
to
be
like
decided,
somewhere
and
and
externalize.
C
Somehow,
because
there's
potentially
a
different
thing
that
has
now
got
to
propagate
that
thing
into
the
asian
cluster
and
is
yet
another
thing
that
has
to
make
the
correct
substitutions
into
the
template.
Given
that
we've
decided
to
put
it
in
cluster
X
and
so
I
I
think
that
that's
what
Torrens
asking
is
is
who
actually
does
each
of
those
steps
and
where
are
the
internet?
How
do
they
communicate
with
each
other
and
how
they
sequence
together
in
the
right
way,
so
the
end
result
is
correct.
A
C
It's
not
actually
in
their
mission
controller.
It's
it's
so
the
way
the
policy
integration
toca
was
done.
It
is
implement
as
an
admission
controller
and
basically
changes
what
user
specified
and
turns
it
into
a
more
specific
said
of
placement
distances
and
then
the
rest
of
the
England
says
Oh
user.
It's
as
if
the
user
chose
that
clusters
they
wanted
to
go
into
rather
than
saying
they
wanted
a
PCI
compliant
cluster,
and
then
you
know
it
does
all
of
the
propagation
and
waiting
and
everything
else,
yeah,
so
I
don't
think
admission.
C
F
Yeah
and
I
think
one
point
that
is
important
to
the
audience
question
is
the
sequencing
of
these
things.
So,
for
example,
one
thought
which
came
after
his
question
is
well
say
we
have
to
enforce
a
placement
directive
and
I
think
that
is
discussed
in
the
document.
Also,
if
it's
not
one
particular
place
when
the
directive
or
object
is
not
tightly
coupled
with
one
particular
object,
a
federated
object,
you
could
have
multiple
of
them.
Also,
then,
who
exactly
takes
this
this
this
fusion
and
at
what?
What
time?
F
B
B
B
F
An
example
that
comes
to
my
mind
is
that,
for
example,
currently,
if
you
see
physician
there
is,
there
is
some
overlap
between
the
scheduling
of
replica
set
based
on
the
weights
or
whatever
annotations
are
applied,
and
something
which
is
fed
back
from
the
system,
for
example,
matrices
in
in
case
of
HPA,
so
they,
this
is
just
an
example
of
two
possibly
shed
in
Polish
policy,
then
forced
on
the
same
object
right
so
I'm
going
I
mean
I
was
thinking
on
that
lines.
What
I'm
saying
is
we
should
probably
avoid
doing
this.
F
B
Mean
the
the
kind
of
pipeline
I
had
in
my
head
was,
you
would
have
you
know
a
template,
you
have
overrides,
you
would
have
placement
and
those
were
kind
of
the
things
that
would
drive
propagation,
and
if
you
have
a
scheduling
mechanism
it
would
actually
influence
either
in
memory
or
or
stored.
You
know
to
NCD
I'm
not
being
specific
about
that,
but
it
would,
you
know,
effectively
change
the
overrides
or
change
the
placement
that
was
required.
C
B
C
C
Well,
get
it
gets
back
to
the
sequencing
stuff
in
that
maybe
I
should
give
it
some
more
thought
before
talking
too
much,
but
I
mean
you
even
have
to
do
substitutions
into
the
template
before
you
do
placement
in
which
clay,
in
which
case
placement
can
be
based
on
template,
substitutions
gory
have
to
do
template
substitutions
after
you
do
placement,
because
maybe,
for
example,
the
template
substitutions
need
input
from
the
placement
decisions
like
which
cluster
the
thing
went
into.
Would.
B
Why
would
that
I
guess
I
guess
if
your
question
is
like
I
can
imagine
like
for
a
given
resource,
it
would
be
a
question
of
like
which
clusters
do
I
go
into
which
to
me
implies
placement
and
then
for
each
cluster.
C
E
B
Although
the
quest
it's
a
good
question
now,
I
mean
the
way
that
we've
been
considering
or
I've
been
considering.
This
anyway
has
been
overrides
or
overrides
if
the
user
specifies
overrides,
but
there's
this
dynamic
scheduling
component,
that's
gonna,
say
set
up
what
account
the
scheduling
value
just
supersedes
these
or
value
by
enabling
scheduling,
they're
implicitly
enabling
behavior
you.
E
B
C
B
B
Yes,
I'm
not
like,
it
definitely
introduces
some
more
layers,
at
least
as
proposed
to
the
existing
model,
but
in
terms
of
like
the
existing
controllers,
they
watch
the
resources
and
hold
them
in
a
local
store
and
then
as
they're
like
watching
for
changes
when
something
changes,
they're
referring
to
the
store
and
the
store,
gets
updated
by
an
informer.
I
think
that
model
is
likely
to
be
consistent
with
what
we
would
implement
for
this
as
well.
C
Yeah
but
it's
possible
because
of
the
way
that,
because
all
of
these
five
things
are
all
inside
the
same
process,
so
they
just
have
an
in-memory
store,
so
I
guess
we
potentially
have
to
externalize
that
somewhere.
If
you
in
your
original
plan,
you
had
this
idea
of
these
things
being
separate
processes
potentially
or
separate
containers,
the
one
doing
the
placement,
the
one
doing,
the
overrides
one
doing
a
propagation
you
find
stood
you
correctly
and
then
how
do
those
like
pass
stuff
between
them?
I
guess
was
the
question.
B
B
B
Mean
someone
really
cared
I,
don't
know,
I'm
saying
this
Paul's
wanted
to
keep
things
as
separate
as
possible,
which
you
know
to
not
preclude
a
user
from
implementing
any
one
of
these
pieces
on
their
own
because
they
have
you
know
a
specific
use
case,
the
require
specialized
behavior
for
placement
or
overrides
or
whatever.
But
it's
not
clear
to
me
how
do
that
cleanly?
Okay.
F
Yeah
before
that,
I
just
wanted
to,
because
I
did
not
get
a
clear
concluded
idea
about
what
we
were
talking
about.
The
problem
is,
let's
take
a
concrete
example
and
take
five
more
minutes
and
maybe
eating
it.
Overlaps
this
example
overlaps
with
the
policy
discussion
that
we
might
have
on
Monday,
so
say
an
external
policy
engine
is
integrated
and
taurine.
F
Correct
me
if
I'm
wrong
from
there
and
the
sequence
of
or
the
flow
is
such
that
the
object
is
created,
a
predicted
object
is
created,
then
a
placement
object
is
created
which
specifies
that
some
portion
or
half
of
that
object
or
the
same
object
goes
into
travel
into
a
circle
both
and
at
the
beginning
there
is
some
policy
applied
which
depends
on
the
property
on
question
one
threads
until
it
is
PCI
compliant
and
based
on
that
policy.
The
placement
that
is
decided
is
everything
goes
to
cluster
one.
F
At
a
later
point
of
time,
this
property
of
clusters,
change
policy.
Engine
or
an
external
entity
gets
to
know
about
this,
so
it
creates
a
new
placement
object
or
a
base
the
same
object,
or
how
does
it
even
get
to
know
that
there
is
an
object
of
a
different
name,
which
is
a
placement
object
which
is
relevant
to
this
particular
federated
object?
F
B
B
If
you're
gonna
apply
policy
to
placement,
how
do
you
do
it?
If
there's
no
one
authoritative
place
to
do
it
for
a
given
resource?
That's
the
problem.
We
need
to
solve.
I.
Think
your
to
your
question
about
whether
scheduling
would
have
the
same
problem.
I
wasn't
anticipating
at
least
until
now,
that
scheduling
would
involve
placements,
and
maybe
you
could
elaborate
as
to
why
that
would
be
the
case
because
I'm
not
sure
it
is
today
so
nature.
You.
F
Know
taking
the
example
of
say,
Topeka
set
itself,
so
if
a
dealing
reference
on
a
replica
set
exchange
such
like,
initially,
it
was
updated
as
50/50
or
one
is
to
one
into
clusters.
Later,
it
is
updated
by
the
user
as
zeros
to
one
so
or
the
weight
is
changed
as
zero
is
to
one
whichever
they
so
one
cluster
now
does
not
get
the
replicas
that
object
at
all,
so
I
think
that
will
be
governed
by
placement
rather
than
there
should
even
thank.
B
I'm
not
sure
it
would
have
to
I
mean
it
would
depend
on
the
propagation
mechanism
if
it
would
implicitly
treat
a
zero
replica
count,
as
do
not
propagate
like
in
the
question,
was
not
selected
or
whether
it
would
continue
to
propagate
with
a
zero
replica
count.
What
does
it
do
today?
I
mean
if
I,
if
the
scheduler
decides
that
a
given
cluster
deserves
zero
replicas
still
topping
eight
centered?
Is
it
knocked
off
data?
If.
F
It's
the
first
time,
that's
happening,
then
it
doesn't
propagate,
but
it
is
a
little
I.
Don't
know
all
it's
stupid
of
it
whatever
or
maybe
it's
some
reason
which
I
don't
clearly
have
it
in
my
mind.
So
if
a
condition
like
this
happens,
that
particular
cluster
should
get
zero,
it's
also
the
replica
set
remains
there
with.
B
B
F
Think
we
already
had
I'm
like
I,
did
try
to
initiate
something
like
this
and
collectively.
Paul
also
said,
this
is
the
same
thing
and
if
rest
of
the
people
also,
we
thought
that
we
will
probably
iterate
for
the
API.
Will
it
create
in
terms
of
code
at
some
location
it
might
be
the
current
current
ripple
or
it
might
be
something
else.
What
we
can
do
right
now
is
maybe,
depending
on
who
all
are
giving
to
other
do
to
contribute.
F
B
I'm,
an
API
server
at
the
API
types
kind
of
mirroring
what
we
have
in
this
document,
I
think.
The
next
step
for
me
is
just
getting
cluster
registry
integrated,
so
we'd
actually
like
run
an
integration
test
with
everything
combined,
because
this,
but
I'm
sort
of
anticipating
in
an
aggregated
model
is
that
we
would
actually
be
aggregating
with
a
queue
API.
So
we're
not
going
to
be
deploying
the
API
as
a
totally
separate
endpoint.
B
B
You
know
my
app
and
I
have
a
replica
set
like
federated
replica
set.
You
know
my
replica
side,
you
know
I'm
replicating
to
the
same
cluster
that
I'm
aggravated
to,
and
it's
actually
going
to
create
a
replica
set
and
not
say
named
space,
but
it's
going
to
be
federated
replica
set,
it's
just
a
template
and
then
a
replica
set
actually
is
running
pods,
which
is
different
from
the
existing
scenario.
B
So
this
is
the
model
suggested
by
like
having
us.
You
know
specific
federated
types
like
unless
we
you
know
reimplemented
as
well,
but
my
goal
is
to
get
that
sort
of
as
a
working
prototype
that
people
can
play
with
sort
of
sorry.
Think
about
like
does
this
make
sense?
Is
this
going
to
be
a
good
user
experience,
or
do
we
have
to
consider
some
other
possibility
of
like
actually
having
a
totally
separate
endpoint
that
has
its
own
binary
cube
objects?
F
Yeah
I
I
did
spend
some
cycles
are
trying
to
to
see
the
current
federated
API
if
we
have
to
maybe
modify
that
and
pull
out
the
cursor
object
and
then
integrate
lesser
registry
with
it.
What
we
might
need
to
do,
I
don't
have
concrete
stuff
yet,
but
if
that
seems
right
like
the
right
direction,
I
can
maybe
try
to
do
that.
I
think.
B
C
Yeah
I
think
the
only
caveat
to
that
is
I
think
we
need
something
sort
of.
We
need
to
start
releasing
stuff
in
a
couple
of
months.
You
know
we
we
need
to
start
making
progress
towards
a
defined
end
point
as
opposed
to
still
be
experimenting
in
three
months
time.
So
I
think.
As
long
as
we
have
some
kind
of
whatever
hanboks
off
on
the
experimentation
phase.