►
From YouTube: Kubernetes Working Group Multitenancy 20190507
Description
May 8, 2019, 11am - Noon (Pacific Time)
To be discussed at this meeting:
- Discuss alternative solutions to the Tenancy CRD (By Yisui Hu, Google)
-- Read them at the end of this doc: https://docs.google.com/document/d/1hpJX5O_siMmNGMvIHvz8Pm7XOjJLz5g57XWrgwWarFw/edit#heading=h.r5a9ppuyrmye
- Discuss the fundamental tenancy concept (KEP?) (By Yisui Hu, Google)
-- https://drive.google.com/open?id=1ddx7UAEPKFPldBh_diksYO4WZXSHDUhm-e6hyNNGYVU
- Propose Virtual Cluster Based Multi-tenancy Solution (By Fei Guo, Alibaba)
-- https://docs.google.com/document/d/1EELeVaduYZ65j4AXg9bp3Kyn38GKDU5fAJ5LFcxt2ZU/edit#heading=h.7tna1yo4dzv (This has been presented to David/Dawn/Daniel)
Notes:
- Multitenancy Repo has a links repo now with all our docs etc in a single place to help with discoverability
https://github.com/kubernetes-sigs/multi-tenancy/blob/master/docs/links.md
A
Okay,
hi
everybody.
This
is
the
multi-tenancy
working
group
for
kubernetes.
It
is
May
7th
Tuesday
and
we
have
a
pretty
packed
agenda.
So
we're
gonna
get
started
a
quick
announcement,
as
you
all
may
have
already
seen
on
the
mailing
list.
My
co-chair
David
is
stepping
down
due
to
having
too
many
commitments
and
not
being
able
to
dedicate
the
time
to
being
the
co-chair.
A
I'm
sure
he'll
continue
to
be
involved
in
the
working
group
since
he's
super
interested
in
multi-tenancy,
but
that
means
that
we
are
going
to
open
the
door
for
people
to
nominate
themselves
to
be
at
the
co-chair
of
the
multi-tenancy
working
group.
So
if
that
is
something
you
are
interested
in,
please
you
know
let
us
know
by
a
mailing
list.
We
are
asking
that
people
self
nominate
and
then
attend
our
next
meeting,
just
to
say
a
couple
words
about
why
you're
interested
and
David
sent
out
an
email.
A
What
sort
of
the
whole
process
based
on
the
working
groups,
Docs
and
everything?
So
take
a
look
at
that
if
you're
interested
definitely
speak
up,
we'd
love
to
have
you
on
board
and
so
I
think
that's
the
main
announcement.
The
other
announcement
is
that
coop
Connie
U
is
in
two
weeks
and
kou
Cuddy
you.
A
We
are
going
to
have
multiple
multiple
sessions
about
multi-tenancy,
we're
going
to
have
an
intro
to
the
working
group,
we're
going
to
have
a
deep
dive
of
some
of
the
work
that's
going
on,
and
we're
also
going
to
have
a
panel
with
several
members
talking
about
multi-tenancy
from
their
perspective.
So
you
know,
pretty
packed
agenda
should
be
some
good
stuff.
Hopefully
we'll
have
some
time
to
socialize
within
attendees
as
well
looking
forward
to
seeing
everyone
there.
A
A
So
I
think
that's
all
the
bookkeeping
announcements
we
have
for
today
definitely
looking
forward
to
seeing
everybody
who
makes
it
out
to
coop
Connie
you
and
if
not,
the
talks
will
all
be
recorded
and
online
for
everyone
to
see
and
with
that
I'm
going
to
hand
it
over
to
East
Way,
who
is
going
to
be
going
over
today,
the
tenancy
CRD
work.
Continuation
of
the
presentation
he
gave
last
week
and
then
fan
Zeng
is
going
to
be
sorry.
Fake
whoa
is
fake.
Well,
one
I
must
have
I'm
sorry,
yeah
I,
see
you.
A
B
Hi
every
is
able
to
see
my
screen,
yep
cool,
so
first
I
have
a
quick
update
on
the
Martin's.
A
repo
I
put
a
link
store
MV
file
in
the
docs
folder,
which
collects
all
the
links
rate
to
the
related
documents,
because
I
hope
you
know,
campaign,
there's
too
many
documents,
scattered
everywhere,
so
I'm
trying
my
best
to
put
all
the
links
together
in
a
single
place.
I
hope
that
will
help
people
to
find
the
relative
of
relevant
documents
and
for
some
proposals.
B
If
people
have
any
idea
whatever
feel
free
to
send
a
PR
to
put
MD
file
in
this
folder
or
if
you
feel
better
with
a
grew
dark,
you
can
also
do
that
right
here
on
Bulldog
and
file
APR
to
put
to
add
a
link
in
this
document
that
will
help
people
to
easily
find
your
document.
Okay.
Now,
let's
move
to
the
tenant
CRD
as
go
to
this
alternative
proposals.
So
after
last
meeting
I
received
a
lot
of
comments
and
feedback
which
are
great
and
I'm
singing
carefully
about
the
tenancy
ID
proposal.
I
demoed.
B
If
we
based
on
assumption
to
tents,
are
restricting
that
namespace,
then
we
cannot
do
that
easily,
and
the
second
problem
is
about
the
ten
names
we
list,
which
is
a
list
inside
a
tenancy
ID,
and
that
brings
in
a
lot
of
start.
People
can
easily
make
mist
with
which
you'll
cause
a
big
penalty.
For
example,
you
have
a
name
namespaces.
B
B
B
So,
based
on
these
three
big
categories
of
problems,
I
came
up
some
alternative
proposals,
so
here
I
put
up
two
proposals
and
I'd
like
to
hear
a
feedback,
or
if
you
have
any
good
ideas,
just
feel
free
to
come
to
raise
your
thoughts
and
I.
Think
it's
good
for
for
him
to
discuss.
So
the
first
one
is
called
tenant,
Adam
administration,
namespace,
so
the
ideal
cry
simple:
we
basically
basically
sink
Rifai.
The
tenancy
ID
to
basically
include
almost
nothing,
probably
only
a
list
of
ten
elements
and
based
on
creation
of
tenancy
Rd.
B
A
special
namespace
it
created
regard
is
how
we
named
that
we
can
figure
out
later
how
to
name
that,
but
a
tenancy
ID
will
will
have
a
corresponding
a
namespace
and
that
names
was
called
intent.
A
demonstration
namespace.
So
in
that
namespace
you
can
create.
You
can
further
creates
additional
CR
DS
for
any
other
purpose,
so
forth.
Hope
we
promote
the
tenant
namespace
structure
as
a
CRE,
so
we're
seeing
the
tent
at
the
ministry
administration
namespace.
We
create
a
namespace
yardie.
B
C
B
Need
some
help
from
the
web
hook
Domitian
webhook,
if
we
want
to
extend
the
permission
for
attend
user
to
create
a
subclass
of
scope
policy
objects
today,
there's
no
quite
reasonable
use
case.
Probably
post
code
policy
is
one
of
them
and
to
use
that
I
you
will
have
to
create
something
in
the
namespace
which
wraps
a
post-game
policy,
and
that
goes
through
some
automation
process
or
the
class
admin
may
review
that
this
is
not
expected.
B
B
B
D
B
Simplify
a
lot
of
work,
we
we
are
doing
with
a
controller
plus
a
at
a
Michigan
web
hook.
Florida
can
help
us
remove
or
eliminate
the
need
of
a
mission
web
hook
and
with
a
generic
pass.
The
entry
gets
more
powerful
to
express
like
a
lot
of
restrictions
applied
on
individual
fields
of
quality
objects
that
basically
makes
it
easier
to
allow
10,
10
users
to
create
classes
called
policies
which
will
be
verified
by
the
policy
energy
to
make
sure
certain
fields
doesn't
doesn't
violate
some
predefined
policy
by
the
cost
enemy,
while
they're
still
being
investigated.
B
There's
a
recent
open
cell
project
called
gatekeeper
which
tries
to
bring
in
the
open
policy
into
criminalists,
and
we
are
trying
to
investigate
that
and
try
to
define
some
trust
model,
some
policies
which
may
be
helpful
for
defining
important
cluster
and
there's
a
community's
positive
working
group
and
will
properly
start
a
discussion
there
and
probably
we
were
working
on
a
proposal,
try
to
find
out
a
few
process
that
could
demonstrate
some
use
case
emergency.
So
if
people
are
interesting,
please
join
this
positive
wing
group
meetings
for
more
in-depth
discussions
about
this.
B
E
One
thing
to
note
is
that
other
multi
tenant
clusters,
like
OpenStack,
do
not
have
a
tenant
administration
area,
so
maybe
we
need
it
specifically
in
cooperate,
is
because
of
cluster
scope,
resources
or
something,
but
that
was
just
kind
of
an
observation
and
also
the
other
point
was
about
the
namespace,
no
accidental
deletion
of
the
list
of
namespaces,
and
things
like
that.
Just
to
understand.
But
carefully
are
we
talking
about
the
case
where,
when
we
do
an
update
of
a
tenancy
Rd,
we
forget
to
include
all
the
namespaces,
and
so
it
accidentally
deletes.
E
Should
we
even
be
providing
a
list
of
namespace
names?
You
know
CRD
to
begin
the
tenancy
re
to
beginning,
because
we
feel
anyway,
going
to
go
down
the
line
of
attaching
the
tenant
name
in
front
of
the
namespace
name.
That
means
that
the
name
that
the
creator
of
the
C
already
provided.
It
is
not
that
important
anymore
because
we
are
making
the
name
unique.
So
should
we
think
about
not
even
asking
for
a
list
of
namespace
names
and
just
providing
the
number
of
namespaces,
and
then
we
ourselves
named
them.
E
As
you
know,
tenant
named
tenant
named
a
you
know:
namespace
one
tenant
name,
a
namespace
two,
so
it's
a
related,
but
slightly
said,
and
and
that
then
can
also
mitigate
this
namespace
deletion.
Accidental
deletion
thing
because
a
you
only
need
to
specify
the
number
of
namespaces,
not
their
names
and
B.
We
can
possibly
put
in
some
protection
against
accidental
deletion
may
or
may
not
need
the
separate
administration
namespace.
C
So
the
big
the
big
issue
I
can
see
with
that
right
is
because
people
are
going
to
put
different
things
in
their
name
spaces
within
their
tenant
and
if
they
needed
the
delete,
namespace
C
and
you
have
a
name
since
ABCDE.
You
know
how
do
you
handle
that
I
guess
deletion
if
you
remove
it,
I
really
liked
Craig's
idea
last
week
or
last
meeting
when
he
mentioned
having
the
namespace
PE
a
separate
custom
resource
that
is
attached
to
a
tenant,
so
the
tenant
CRD
doesn't
have
any
namespace
attached
to
it.
B
B
So
so
original
I
was
thinking
like
moves
like
a
like
move,
the
ten
teams
racing
to
a
separate
class
scope,
CRD
and
actually
there's
further
use.
Kid
comes
up.
I
think
they
will
demonstrate
later
about
a
leader
by
his
case,
there's
other
things
similar
to
like
similar
to
ten
namespace
ERD.
There's
more
things
want
to
be
fair
to
a
tenant,
so
if
we
put
them
on
as
a
classical
piane,
probably
fine
I
mentioned
this.
B
The
Machine
namespace
basically
gives
tend
adding
more
flexibility
that
they
can
create,
create
something
that,
in
the
content,
without
a
more
broad
and
publishing
on
the
classical,
because
it's
very
difficult,
if
it's
three
difficult
grant
tent
users
with
certain
finishing
on
class
scope,
but
only
limited
to
a
set
of
resources,
and
that's
actually
so
the
prom
is
actually,
if
we
grant
the
permission
to
the
tent
users
on
a
subset
of
resources.
My
name's,
the
Macan
basically
and
work
butter
is
very
communicated
and
it
doesn't
have
a
good
user
experience.
B
B
E
So
this
kind
of
like
several
threads
here
that
each
need
to
be
kind
of
discussed
here
right.
One
is
the
accidental
deletion
problem,
which
my
opinion
is
type
passed
lately
to
the
naming
problem
as
well,
which
we
haven't
mentioned
and
then
there's
the
cluster
scope,
CRD
creation.
What
would
ingress
be
another
example
of
a
cluster
scope,
theory
that
we
may
want
tenants
to
kind
of
control.
E
B
Let
me
sing
I
I,
don't
have
concrete
use
case
in
my
mind
today,
right
now,
so
I'm
not
sure
I
can
probably
dig
deeper
and
find
out
some
use
case
and
one
one
other
things
I
I
remember
is
I
was
thinking
about
the
princess
and
volumes
and
story
class
I
think
it's
quite
reasonable
for
all
these
to
be
managed
by
cast
aiming
but
I'm,
not
sure.
If,
if
in
some
use
cases
the
tenth
users
will
probably
need
it
to
define
some
story
class.
B
B
C
So
I
guess
I
mean
one
question,
or
one
thing
that
we
should
take
is
I.
Guess
an
action
is
to
reach
out
to
I
guess
how
the
I
don't
know
would
be
Sega
architecture
or
API
machinery,
maybe
just
thoughts
that
they've
had
on
other
people
with
these
issues,
because
this
can't
be
the
only
ones
that
are
trying
to
put
permissions
around
cluster
or
scoped
items
but
allow
them
in
certain
name.
Spacing
I,
don't
know.
C
For
the
users
right
so
is
there
work
being
done
to
improve
that,
or
is
this
basically
the
lay
of
the
land
and
we
have
to?
We
have
to
be
so
I
guess
the
question
is:
do
we
need
to
solve
this
or
is
it
being
worked
on
and
we
can
basically
work
in
tandem
to
have
an
eventual
solution
that
meets
our
needs
so.
B
I'm
going
to
talk
about
the
tenant
concept
in
communities
which
has
some
sauce
on
that
so
I
have
feeling
like.
If
we
write
on
existing
mechanisms
in
communities,
we
may
come
up
with
a
proposal,
but
I
seem
like
it's
like
with
some
sort
of
like
hacking
or
a
pension,
whatever
it
doesn't
come
up
a
very
elegant,
more
transient
solution.
E
Go
ahead,
complete
your
what
you
were
saying
please
so
I'm
done
so
I
mean
yeah.
One
additional
comment
I
wanted
to
make
was
if
we
do
not
give
any
tenant
user
or
admin
the
ability
to
create
cluster
scope
resources-
initially,
maybe
that's.
Okay.
Maybe
we
differ
that
as
something
that
only
a
cluster
admin
can
do
on
behalf
of
a
tenant,
so
it
may
not
be
as
user
friendly
as
some
people
might
like,
but
we
don't
need
to
create
functionality
before
we
truly
need
it.
E
So
we
could
say
that,
for
now
there
has
to
be
an
interlock
between
the
tenant
admin
in
the
cluster
admin
when
the
tenant
needs.
They
trust
us
copious
resource
created
on
his
behalf,
I'm
just
kind
of
just
coming
out
a
possibility
here
and
right,
and
then
we
wait
to
see
what
the
problem
is
truly
necessary
before
we
curved
a
solution
for
yeah.
B
That's
a
that's
a
that's
a
fair
point.
The
only
concern
I
have
is
to
so.
Basically
we
fall
to
the
extendibility
problem,
so
we
do
need
some
externality
to
allow
other
visual
to
be
plugging
with
tenants.
So
without
a
tent
at
the
ministry
names
list,
we
are
currently
finding
or
other
series
as
class
scope
and
I'm,
not
sure
that
will
be
critical.
B
Some
other
becames
to
convert
all
these
ERU
into
names
base
code
probably
would
be
a
problem
in
future,
but
I
agree.
We
can
stay
with
a
relatively
simple
mckennon
today
when
the
prom
comes
up,
we
can
try
to
solve
later
so
I
even
more
comments,
because
I
need
to
be
fast.
I
was
switched
to
lieutenant
concepts
in
kubernetes.
This
documents
a
bit
basically
a
bit
abstract
and
talking
about
the
source
about
why
we
need
a
tenant,
conserving
chameleons.
What
helps
in
basically
how?
B
B
B
So
that
means
we
can
find
the
tenant
by
using
a
use
identity,
but
actually
that
brings
a
lot
of
problems.
The
biggest
one
is
basically
is
MBM
viewers,
because
a
single
you
that
entity,
a
single
user,
can
belong
to
multiple
tenants
when
making
a
API
call
without
explicitly
specify
the
tender
name,
it
will
be
difficult
to
figure
out
which
tender
you
that
represents,
and
that's
also
extent
to
the
problem
with
service
accounts.
Service
counts
is
more
like
a
system
level
and
more
usual
to
tense.
B
So
when
a
control
is
running,
how
can
you
tell
what
you
tenants
this
control?
It
represents
in
terms
of
a
specific
API
call,
so
a
few
ways
to
work
around
this.
We
can
introduce
a
tenth
name
in
IKEA.
Pass
is
a
high-level
scope
above
namespace
definitely
does
require
significant
change
of
APs
and
a
lot
of
complication
being
implied
by
this
and
the
alternatives.
Big.
We
can
introduce
a
new
edge,
giving
header
like
X
canary
tenant
and
put
any
name
there.
I
wish
also
replace
the
tenant
must
be
explicitly
specified
for
every
API
request.
B
B
So
if
we
associated
hand
on
the
target
resource,
regardless
of
who's,
making
the
call
that
may
simplify
the
problem
and
that
can
eliminate
his
the
the
ambiguity
so
a
very
simple
way:
we
don't
need
any
change
on
the
Canaries.
We
just
put
a
label
on
the
resource
and
for
every
key
call,
we
figure
out
the
label
from
targeted
resource,
and
that
represents
the
name
of
the
tenant
for
all
the
names.
Basically,
resources.
We
just
put
a
label
on
namespace.
Don't
you
take
a
deep
deep
into
every
resource?
B
So
that's
actually
a
better
way
to
represent
the
tenant,
to
pick
out
a
name
for
each
of
a
command
eg
a
call.
So
this
is
basically
two
ways
to
define
a
tenant,
and
the
last
thing
I
want
to
mention
is
I.
Call
it
impractical
sorts,
but
I'm,
not
sure
it's
reading,
right
or
not
so
I.
Remember,
there's
a
few
discussion
happened
quite
long
ago
about
the
hierarchical
namespace
to
some
buttons.
It
basically
hierarchical.
Namespace
is
a
lot
of
weight,
better
solution
which
texts
a
lot
of
problems
we
added
causes
all
today.
B
So
first
like
naming
conflict
with
hierarchical
namespace,
we
don't
have
any
conflict
and
the
result
will
availability.
We
basically
can
encapsulate
encapsulated
resources
within
a
container
which
is
like
a
cup
a
tenant
without,
like
a
single
list,
call
will
list
all
the
classes
go
to
resources
regards
which
they
belong
to,
and
it's
also
possible
to
introduce
like
a
nest.
B
Lieutenant
though
it's
not
quite
practical,
to
have
a
very
deep,
deep
nasty
level
of
tenants,
but
usually
like
a
10
plus
a
sub
tenant
will
be
quite
common
in
a
lot
of
use
cases
and
the
last
one
it
can
be
quite
useful
is
for
the
policy
inheritance
and
override
today
is
really
difficult
to
do
a
deposit,
inheritance
and
overwrite.
All
the
policy
objects
are
highlighted,
one
task,
scope
or
namespace
scope
and
there's
no
algorithm
fitting
them
well.
B
F
Can
you
hear
me?
Yes,
please
yep,
so
in
the
previous
section
that
you
were
mentioned,
that
the
tenant
will
be
represented
as
the
label
if
I
ever
work
load
right,
for
example,
Leah
pod,
and
it
has
to
do
so-
use
particular
CPU
or
memory
so
have
a
restriction
for
my
tenant.
So
how
using
the
labels
I
have
to
derive
the
namespaces
first
right
so
currently
in
the
POC,
the
namespace
is
derived
from
the
tenant
name
and
the
name
itself,
so
it
will
be
reflected
if
I
have
a
workload.
F
D
F
For
the
workload
actually
fare,
workload
right
I
have
to
specifically
have
to
say
a
tenant
in
s1,
or
something
like
that.
So
in
case,
if
you
have
a
label
right,
there
should
be
a
mechanism
or
an
admission.
You
take
muting
Tabuk
that
will
insert
get
the
tenant
and
the
namespace,
and
it
has
to
insert
that
civic
name
space
for
each
pod
or
the
workloads
right.
He
said
yeah
mechanism.
We
have
to
do
that.
Yeah.
B
F
D
C
B
Like
a
little
Oh
similar
to
introduce
this
hierarchical
namespace,
if
we
abstract
the
namespace
using
like
a
pass
format
using
a
stash
to
split
that,
basically,
you
can
see
that
tenant
name
becomes
a
prefix
of
namespaces,
okay,
okay,
thank
you,
yeah!
So
I
I
like
to
dig
a
lot
deeper
bout,
the
hierarchical
namespace
to
try
to
gods,
also
more
details.
G
B
G
B
So
there's
a
few
ways
to
introduce
a
turns
into
natively
into
kubernetes.
Quite
an
intuitive
way
is
if
we
think
tenants
is
a
layer
above
namespaces,
so
today's
api
server,
it
has
a
epa
passed
like
a
psst,
mustache,
namespaces
and
whatever
names
to
name
and
then
stereotype,
/resource
name
so
by
introduced
at
nth.
One
possible
way
is
like
regions.
Api
tasks
like
api's,
meanwhile
slash
tenant
selection.
They
aim
then.
G
B
G
B
G
B
G
And
but
what
will
be
the
concept
called
like
I
mean?
Basically,
it
will
be
somehow
like
using
a
label
Club
them
together
or
using
a
top-level
object
that
will
basically
somehow
specify
all
the
users
or
service
accounts
that
in
composite
I,
mean
I,
know
I'm,
asking
details
that
you
probably
still
thinking
I'm,
just
asking
what
what's
the
thinking
there
yeah.
B
So
that's
where
basically,
the
the
design
details
or
implementing
details.
There
are
few
ways
so
one
way
we
can
have
a
ceoddi
call
the
tenant
and
include
all
these
users
and
groups
and
service
counts
and
other
ways
like
labels
is
also
possible.
But
a
label
has
other
problems
that
can
I'm
not
sure,
maybe
that
problem
that
would
be
a
little
difficult
to
match
to
associate
the
users
because
user
doesn't
have
objects
in
select
canaries.
So
you
can.
C
E
Right
so
one
way
to
look
at
it
is
that
really
there
there
are
two
at
a
high
level.
We
are
kind
of
talking
about
two
categories
of
solutions:
one
is
a
tenant,
is
a
CRD
sitting
on
top
of
a
collection
of
namespaces
and
resources
and
users
are
associated
with
that
content,
indirectly,
not
directly
because,
as
you
see
said,
there's
no
users
and
communities,
and
then
the
second
thing
is
the
hierarchical
namespaces.
E
Now
it
seems
that
correct
me
if
I'm
wrong,
you
say,
but
if
we
do
hierarchical
namespaces,
that's
really
going
back
to
the
API
server
team
and
asking
them
to
extend
the
namespace
model
itself
to
create
hierarchical,
namespaces
and
just
handle
it
that
way,
and
then
what
what
we've
been
in
this
team
sort
of
talking
about
and
prototyping
is
more
like
the
concept
of
a
tenancy
already
sitting
on
top
of
a
collection
of
namespaces.
Yes,.
B
D
B
And
also
have
a
lot
of
problem,
not
quite
well
solved
with
the
hierarchical
namespace
is
I
just
have
to
just
rough
idea
and
I'm
current
doing
some
investigation.
So
basically
I'm
not
sure
this
is
practical
or
in
practice,
so
I'll
call
that
in
practical
I
will
probably
get
more
data.
Almost
I
will
probably
be
more
clear
in
a
few
weeks,
so
I'd
like
to
have
Plan
Update
in
next
meeting.
B
G
To
clarify
my
thinking
again,
so
it's
the
purpose
of
this
working
group
is
to
come
up
with
this
list
of
proposals
or
or
suggestions
on.
How
we
can
implement
tenancy
in
cabinet
is
right
and,
and
it
could
be
in
the
form
of
a
yadi,
it
could
be
changes
to
coal
kubernetes,
but
it
could
be
either,
but
we,
the
purpose,
is
just
to
figure
out.
G
A
So
we
have
a
project
plan
that
is
included
in
East
Way's
list
of
links
that
you
can
review
to
kind
of
see
what
our
proposal
is
for
how
to
move
tenancy
forward.
We've
broken
it
down
into
a
couple
different
milestones
that
we'd
like
to
achieve
and
gone
over
and
a
couple
other
meetings
that
have
all
been
recorded,
so
I
would
definitely
encourage
you
to
take
a
look
at
that
and
then
ask
any
questions
on
the
dock.
Thanks.
G
G
E
G
A
G
G
Question
is
a
long
back.
I
had
created
an
issue
related
to
how
persistent
volume
should
be
multi-tenant
as
well
like
they're,
currently,
cluster
scope.
Is
there
anybody
thinking
about?
How
would
that
work
in
this
multi-tenancy
solution?
If
I
wanted
permissions
on
persistent
volumes,
which
are
right
now,
clusters
code,
I.
B
E
A
F
H
Okay,
so
thanks
everybody,
my
name
is
Fay
I'm
coming
from
Mario
Bava,
so
I'd
like
to
take
this
opportunity
to
introduce
a
kind
of
relatively
another
ended.
You
solve
the
monotonicity
wrong,
because
this
is
associate
a
very
strictly
for
about
a
used
case.
It
won't
take
long.
Oh,
is
this
a
kind
of
very
long
talk,
but
today
I'm
going
to
spend
about
ten
minutes
to
just
go
over
introduction
and
let's
have
a
deep
discussion
through
the
dogs
or
even
any
channels
in
the
working,
the
working
groups.
Okay,.
H
How
about
this
yep?
Sorry
all
right!
So
so
so
we
we
come
up
with
proposal
kind
of
tied
you
a
lot
about,
because
K
speakers,
you
know
Ali
Baba.
We
we
want
to
share
a
lot
of
good
things
for
multi-tenancy.
The
my
purpose
is
for
better
customer
ID
ability
and
a
research
in
relation.
Where
are
you
Ali
Baba,
our
talent?
It
beats
typically
in
general,
bu
has
some
kind
of
special
requirements.
First,
they
want
a
harder
multi-tenancy
model
and
they
they
have.
H
A
special
requirement
is
saying
that,
oh,
please,
don't
use
namespace
you
group
resource,
because
many
teams
has
already
used
namespace
to
organize
the
video
application
it
reintroduce
a
tenant
model
that
has
to
use
namespace,
which
means
many
team
has
to
change
the
way
that
they
deploy.
That
application,
which
is
not
preferred
so
because
of
such
use
case.
We
propose,
or
we
call
virtual
class
based
multi-tenancy
solution.
So
some
key
some
key
factors
about
the
widget
cluster.
H
Basically,
there
was
no
modification
required
on
the
control
plane,
the
core
control
plane
and
the
last
Tony,
because
we
using
rotary
cars,
we
allowed
multiple
tenants
JAC,
we
share
the
same
fiscal
notes
is
a
unified
scheduler
provided
by
the
pseudo
master.
Now
we
this
our
character,
we
can
achieve
kind
of
high
resources
for
the
super
master
I
just
compared
with,
because
in
this
wasn't
group
the
user
has
mentioned
a
lot
of
times
by
the
efforts
for
the
using
the
namespace
to
group.
H
The
resource
prevent
and
a
prison
has
a
talent
and
they
use
the
are
back.
You
provide
the
isolation,
I
don't
want.
You
make
a
few
comparisons
between
the
current
approach
estimation
and
the
approach
to
that.
We
propose.
First
I
want
to
mention
that
we,
these
two
proposed
actors
share,
share
the
same
concept,
telling
the
concept,
so
we
we
both
so
so.
H
Basically,
what
you
cross
that
we
still
treat
our
tenant
is
a
kind
of
a
group
of
the
resource
we,
the
group
tenant,
is
a
kind
of
a
body
and
isolated
amount
of
computer
storage,
networking
and
the
controlling
resource
busy,
nor
kubernetes
cluster
and
the
in
in
the
namespace
group
approach,
which
is
the
approach
you
summation.
The
tenants
share
the
same
control
playing
well
in
which
a
cluster
we
provides
a
dedicated
control
crane
to
each
tenant,
which
kids
come
at
a
cost
of
additional
resource
for
the
running
at
helena
master
like
the
tan
and
a.
H
G
H
So
in
namespace
groups,
the
tenant
resource
organized
organized
by
namespace
in
what
you
cluster
general
resource,
are
presented
as
rear
kubernetes
cluster
is
a
for
ApS
course,
and
the
namespace
group
relies
on
our
back
to
provide
API
level
isolation,
but
the
virtual
class
are
naturally
provide
a
strong,
API
lower
isolation
because
it
has
his
own
data.
Katie
come
from
playing
in
this
approach.
There
is
a
common
question,
which
is
a
common
dream
rise,
because
the
motocrosser
needs
another
tether.
The
master
which
is
is
doesn't
come
for
free,
so
we
agree.
This
is
alleged.
H
This
is
a
very
valid
concept.
Okay,
if
the
tenants
only
needs
small
model
resource
and
around
few
parts,
maybe
running
a
full
set
of
the
counter
plane
is
too
costly.
Well,
indeed,
I
believe
there
are
some
use
cases
like
what
we
have
in
our
abhava,
where
each
tenant
actually
needs
to
consume
large
amount
of
resource,
we're
only
a
lot
of
parts,
then
the
cost
of
the
rani
another
control
playing
is
kind
of
mitigated
by
the
factor
of
a
large
amount
of
the
actual
workload
pods.
H
Another
thing
is:
this
is
a
the
approach
that
we
are
talking
about.
Having
another
control
plane
is,
a
very
club
is
a
classic
classic.
You
know
coop
on
coop
solution,
I
think
this.
This
solution
has
been
using
many
places
and
we
have
quite
a
few
ways
to
actually
reduce
to
the
cost
of
the
control
plane,
like
you
know,
shared
a
180,
CD
or
heavy
less
replicas
eccentric.
So,
overall
we
see
this
approach
is
a
supplemental
solution
to
the
namespace
group
to
deal
with
different
use
cases
and
I.
H
So
next,
I
will
briefly
describe
the
architecture
and
I
will
skip
the
details
and
for
the
readers,
if
you,
if
you
are
interested,
you
can
see
the
rest
of
the
talk
and
we
also
present
provide
more
detail.
The
links
for
the
slides
through
presentation,
if
you're
interested
so
into
this
medium,
just
quickly
go
over
the
high-level
architecture.
H
So
basically,
if
you
look
at
this
pitch
a
picture,
so
we
have
a
super
master
which
is
which
is
come
in
in
a
rare
color
which
which
is
actually
the
resource
provider
which
managing
all
the
physical
resources,
which
means
in
each
physical
knows
the
koobideh
stocks.
Your
super
master
owning
the
cuckoo,
but
it
doesn't
come
to
you
any
notes
on
the
telling
the
master
and
the
tenor
the
master
from
the
tenant
perspective.
They
are.
You
know
I
modified
the
kubernetes
control
plane
where
in
which
each
nose
is
actually
weary.
H
Cause
were
show
note
which
is
just
simply
or
what
you
which
is
a
you
know:
api
object,
maps
to
or
work
the
you
know
the
agency
each
note
not
Maps
to
de-ice
couplets,
so
the
V&A
Giunta
stands
as
a
as
a
proxy
to
forward
all
the
you
know:
api
calls
from
the
tenant
of
master
which
caused
the
couplet
they
afford
all
the
api
calls
to
the
kabbalists
and
do
some
manipulation
based
on
the
tenant
information
and
the
to
actual
complete
call
and
return
the
results
for
the
each
tenant,
a
master
economy.
By
some.
H
You
know
manipulator
of
the
results
list
and
and
the
tenant
controller,
which
is
new
components
which
needs
to
be
taking
care
of
the
tenant
creation
and
also
some
mechanism
to
syncing
to
to
sync
all
the
objects
from
the
Tanana
master
to
super
master
and-
and
there
is
a
unified
scheduler,
which
is
kind
of
odd,
apart
scheduling.
So
to
avoid
the
order,
you
know
split,
brain
problem
if
you
have
multiple
clusters,
multiple
schedules,
so
so
so.
H
Lastly,
so
or
all
the
you
know,
mappings
big
reach,
which
so
all
the
objects
that
are
created
for
that
tenants
and
to
that
a
tenant
map
is
he's
kept
his
his
stars
in
the
talent
of
ten
an
object
in
the
super
master.
So
I
give
you
a
single
example.
Maybe
we
can
understand
this
actor,
be
better
just
examples
now
with
this
architecture,
although
we
didn't
change
the
core
Cobras
come
component,
but
the
many
workflow
as
the
affected.
Let's
see
how
in
this
architecture,
if
your
tenant,
why
don't
you
create
a
part?
H
Now
the
workflow
becomes
the
tenant
first
created
a
passing,
a
tender
master,
then
the
sync
manager
which
stays
inside
at
the
tenant
controller.
We
are
copied,
apart
from
a
tenant
of
Master
G
super
master
and
as
tenant,
perfect
stupa
name
label.
Key
very
sanity
avoided
complex.
Then
the
unified
scalar
choose
a
node
in
the
super
mustard
for
the
path.
Then
the
Kunitz
create
a
pause
Innes
in
the
second
node,
and
the
sync
manager
needs
to
true
another
object
in
a
tenant
master
which
used
the
noses
of
via
agent
at
one
time
info.
H
C
H
No
actually,
so,
if
you
look
at
this
way,
the
talent
is
is
actually
a
path,
a
pod
irani,
another
API
server.
That's
it!
You
don't
need
to
do
so
form.
The
super
master
perspective
so
definitely
need
to
register
to
the
tenant
controller,
because
tenant
controller
needed
great
Rose
tenant
a
master
but
from
the
CMS
perspective
is
just
a
CRE
is
just
a
pause
so.
H
Then
the
master
has
a
controller,
but
he
doesn't
has
a
scheduler
or
the
poor
scheduling
is
done
by
the
super
master.
So
thus
the
scheduler
is
done
because,
as
I
said
in
the
workflow
every
time,
if
you
want
to
do
something,
if
the
tenant
I
want
to
do
something
in
a
tenant
master,
we
have
a
controller
to
copy
the
Intel
pass
into
the
super
master
etcd.
Now
the
super
master
knows
you
want
to
do
something
and
for
the
full
scheduling
super
master,
we're
due
scheduled
for
you
are
using
the
unified
scheduler,
so
the
tenant.
H
So
all
the
scheduling
party
hidden
from
the
tanana
master
perspective.
So
in
the
end,
as
I
said
from
Canon
master
view,
they
see
a
nose,
which
is
the
word
sure,
nose
and
the
nose
the
nose
the
mapping
the
map.
He
knows,
information
is
not
accurate.
Information
is
only
an
agent
information,
so
this
is
a
week.
This
is
a
horizontal
we
do.
Is
we
have
another
abstraction
about
a
nose?
How
do
you.
H
Is
no
M,
no,
that
you
lieutenant
in
this
model
so
every
time
so
this,
so
this
is
kind
of
these
cases
from
the
book
format
and
a
perspective
that
you
just
need
you,
how
create
a
pass?
That's
it!
So
if
the
past
needs
to
decode
shoot,
if
it
knows
a
new
version
of
you
have
a
pop-up
in
the
end,
you
know
it's
kind
of
automatically
scaling
and.
G
H
G
D
H
C
Sure
so
I
guess
so.
If
I
wanted
to
create
a
new
storage
class
and
deploy,
you
know
Brooks,
F
or
whatever,
and
it's
going
to
basically
have
repercussions
on
all
of
the
worker
nodes
and
basically
take
up
disk
on
all
the
nodes.
I
guess
how
is
that
handled
and
if
10
a
antenna
be
needed
it?
You
know
there
could
be
potential
conflicts,
I'm
just
used
to
genuine
storage.
H
Across
is
the
one
thing
that
I
want.
You
I
think
this
is
the
one
thing
that
we
should
is
allow
people
to
do
in
this
model,
because
the
biggest
problem
is
if
the
people
create
some
so
like,
like
a
precision
volume,
probably
to
concede
warning
me
wheezing
this,
a
master,
the
unified
scheduler
is
not
a
well
for
that.
That's
a
problem
so,
but.
D
All
right,
this
is
young
from
Alibaba.
So
if
we
won't
solve
this
problem,
we
need
to
make
the
attendant
controller
aware
of
all
the
pendant
master
like
resources,
so
it
can
copy
to
the
super
master.
So
we
have
the
appealing
to
a
wider
conflict,
because
at
least
a
why
the
naming
conflict,
but
if
it
is
a
actual
resource
conflict,
we
cannot
really
avoid
it.
Otherwise,
we
need
to
say:
okay,
this
tendon
will
have
some
delicate
nodes,
and
only
those
that
can
note
will
have
this
cannot
against
all
our.
C
H
H
Yeah,
so
this
is,
however,
because
part
of
an
overview
of
the
project,
so
I
welcome
any
comments.
So
please
just
write
a
loyal
comments
in
a
talk,
so
we'd
like
to
have
further
discussion.
If
you
guys
are
interested,
we
can
come,
we
can
use
one
of
the
following
meeting
for
further
fortune.
That
is,
a
people
wants
to
give
got
there.
So
yeah.