►
From YouTube: Kubernetes Resource Management WG 20171025
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
All
right,
so
this
is
the
October
25th
meeting
of
the
Christmas
magic
workgroup
bang.
The
hell
out
realizes
I
need
realize,
need
to
wish
my
mother
happy
birthday,
but
I
got
three
topics
on
the
agenda:
one
which
I
threw
on
there
about
quota
vise,
plugins
and
I.
Guess
the
stub.
So
let's
jump
right
into
it.
So.
A
I'm
happy
you
joined
so
basically
I
look
at
quota
right
now.
It
doesn't
really
have
a
home
in
any
particular
sake,
and
historically
I
had
done
some
work
there,
but
then
it
like
predated
the
sig
model
and
right
now
it
seems
like
when
we
try
to
do
resource
management.
Related
topics
feels
to
me
like
quotas,
a
hodgepodge
of
maybe
API
machinery
and
scheduling,
and
this
group
so
I
wanted
to
use
this
forum
as
like
a
vehicle
to
discuss
some.
A
What,
if
stuff
that
was
going
on
so
people
are
aware
and
if
they
want
to
engage
and
help
like
it
would
be
appreciated
and
then
to
settle
on
some
of
the
discussions
for
the
priority
and
preemption
is
on
earth
quota
and
then
use
this
as
a
case
study
for
what
we
want
to
do
with
the
resource.
Api
moving
forward,
so
I'm
not
sure
if
Harry
or
I
know
Connor
you're
here
I
know,
there's
been
some
discussion
around
people
being
reinter
estill
resource
classes,
but
I
think
there
was
some
requisite
legwork.
A
B
A
Well,
I'm,
sorry
about
that.
I
feel
like
Seth,
who
had
the
similar
problems
are
here
in
the
day
or
yesterday.
I,
don't
know
if
I'm
gonna
be
able
to
trick
zoom
into
letting
me
share
my
window
here.
So
let
me
do
a
quick
heads-up.
So
basically,
I
spent
some
time
last
few
days,
rewriting
Florida
to
allow
you
to
do
photo
on
any
cube
resource
via
Discovery,
but
I
don't
have
any
ability
after
this
PR
lands,
hopefully
soon
to
do
quota
on
custom
resource
definitions.
A
Crd
stuff,
aside
from
that
there,
like
I,
was
potentially
going
to
use
this
time
to
going
update
quota
to
support
all
of
the
new
compute
resources
that
we've
been
talking
about
at
least
huge
pages
and
maybe
devices
or
ephemeral,
storage
or
any
of
that
stuff,
and
one
sitting
back
to
thinking
about
doing
that.
One
of
the
things
that
would
be
useful
is
to
know
like
what
we
imagine
like
the
scope
of
a
future
resource.
A
Api
should
be
so
like
I
could
know,
for
example,
that
if
a
particular
resource
is
over
committable
and
it's
tied
to
pogs
like
give
me,
the
list
of
resources
like
to
be
supported
there
or
similarly,
actually
persistent
buying
things.
I
would
like
some
way
of
discovering
what
resources
could
be
associated
with
those
I.
Don't
know,
there's
if
you
had
had
thoughts
on.
A
B
B
There
are
few
use
cases
that
I
was
considering
for
a
better
resource.
Api,
that's
not
really
an
exhaustive
list,
but
some
of
the
ones
that
one
would
be
just
discovering
what
resources
exist
within
the
cluster.
A
lot
of
these
resources
might
might
be
tied
to
a
node,
but
some
of
them
might
be
cluster
level
resources.
So
just
like
this
cuz
already
is
like
one
aspect
now
that
aspect
complete
like
having.
B
D
B
A
A
Api
like
I,
was
game
for
hearing
it,
because
I'm
tired
of
working
on
quota,
sometimes
and
if
I,
could
support
like
all
the
pod
resources
relatively
quickly
like
I'd,
want
to
do
it
now,
while
I
was
looking
at
it
and
if
I
can't
do
it
now,
because
people
don't
have
like
a
good
scheme
in
mind,
then
I
just
want
to
be
able
to
enumerate
that
and
then
queue
up
prioritizing
doing
the
resource.
Api
say
faster
than
doing
some
of
the
other
use
cases
that
we've
discussed.
A
B
If
you
are
extending
the
API,
like
I,
mean
if
you
wanna,
keep
existing
string
format,
which
is
like
resource
dollar
a
little
bit
odd
resource
name,
then
I
guess
like
API
changes
and
really
that
big.
But
if
you
are
considering
like
adding
to
represent
external
resources,
then
I
think
it's
probably
worth
passion
about
how
the
resource
API
would
look
like
and
and
and
see
if
there
is
a
forward
compatibility
path
rather
than
like
annex
something
in
a
short-term.
B
B
So,
like
people
expect
that
to
just
work,
I
mean,
even
though
we
don't
actually
advertise
that
that's
what
they
sort
of
expect.
Looking
at
the
issues
that
people
have
filed
around
quota,
so
I
think
that
will
not
view
the
behavior,
except
that
resource
names
don't
have
like
ridiculously
long
strength.
B
A
So
I
guess
my
takeaway
is
like
there's
no
obvious
answer
right
now
for
folks
here
and
so.
If,
in
the
1:9
release,
we
grow
quota
to
at
least
be
able
to
quote
all
the
standard
things,
that's
a
nice
move,
and
hopefully
if
people
could
look
at
that
PR
and
say
the
other
seems
generally
sensible
from
an
API
standpoint
that
be
appreciated
and
then
the
other
topic
then
related
and
then
I
guess.
A
I
could
pause
on
how
to
manage
pod
related
compute
resources
more
in
quota
until
the
iron
out
the
resource
API
more
hopefully
like
a
one
time
time
frame,
the
Ellaria
I
did
want
to
be
able
to
discuss
a
little
bit
now
following
David
and
folks
was
just
the
doing
quota
on
priority.
So
I
did
try
to
comment
on
that
PR
design
with
some
thoughts
on
on
tweaks
I,
don't
know
about
it
yet
chance
to
review
that.
Or
did
you
have
any
input.
E
A
A
E
A
E
A
E
F
A
To
me
so
wish
I'm!
Okay
with
that,
because,
like
users
can
list
all
storage
classes
now
right,
so
like
the
idea
that
they
could
list
all
priority
classes
and
nuts.
To
me,
the
problem
I'm
having
associating
was
like
if
the
mental
mapping
between
priority
class
name
to
resolve
priority
value
is
something
that
is
difficult,
so
difficult
for
users
to
follow.
That,
like.
A
E
E
Anyhow,
so
I
hear
you
Derek,
but
I
I
would
say
that
this
is
pro.
So
imagine
that
you
are
a
user.
You
specify
a
priority
class
name,
let's
say
medium
priority.
This
is
actual
name
medium
priority
in
your
pot
spec.
And
then,
when
you
try
to
run
this
pot,
you
will
get
a
message
saying
that
you
don't
have
enough
koda
in
range
100
to
1000.
So
it's
like
that's
fine,
I
guess
would
let's
say
the.
C
C
Then
it's
they're
having
to
think
about
the
numbers,
and
one
of
the
reasons
for
having
the
name
separate
from
the
number
was
so
like
an
administrator
could
change
the
mapping
later
or
whatever.
So
that
seems
like
a
downside
of
trying
to
do
the
numeric
ranges,
because
that
was
not,
and
the
users
aren't
supposed
to
think
about
the
numbers.
C
G
E
A
So,
like
I,
just
pasted
a
link
to
a
comment,
I
threw
on
the
latest
design
like
I,
think
for
at
least
how
I
write
had
I
imagined.
We
would
use
priority
in
our
hosted
clusters.
This
would
let
me
configure
our
clusters
in
a
way
that,
like
end
users
can't
get
the
same
priority
as
our
cluster
control
point
services
or
like
what
you
did
today
with
critical
pods.
So
if,
if
people
are
okay
with
this
basic
way
of
configuring,
the
system
to
restrict
bounded
quota
I'm
down
to
quota
for
particular
priorities
and
I-
guess
I'm
I'm.
A
G
G
F
E
A
So
if
folks
are
really
excited
about
custom
resources
and
being
able
to
manage
those
I'd
love
to
sync
up
with
you
to
figure
out
how
I
can
then
get
this
under
quota,
and
then
this
shall
hold
on
managing
more
populated
compute
resources
that
are
not
standard
resources
under
photos,
till
1:10
and
then
I
think
assuming
David
and
Bala.
You
guys
are
fine
with
my
last
comment
on
that
PR
I
think
you
can
unblock
you
guys
from
doing
I
already
work
and
I'm
sorry
for
being
slow
on
that
yeah.
B
A
Since
I
can't
share
my
screen,
one
of
the
things
I
want
to
do
like
if
we
all
agreed
on
that
was
get
that
slotted
in
this
roadmap
document
and
since
I
can't
share
my
screen
in
assuming
there's
no
disagreement
like
right.
Now
we
live
resource
classes.
Api
is
like
a
a
110
target
like
I.
Just
want
to
update
that
to
express
that
we
think
before
we
would
group
resources.
I
want
to
go
to
some
with
discovery
of
resources.
That
would
you
about
the
like
priority
queue.
G
A
B
A
A
B
I
A
J
G
J
A
architecture
proposal
and
that
I
think
we've
been
we've
been
missing
for
the
past
few
weeks.
So
basically
the
first
part
is
about
numbering
and
numbers
and
basically,
what
bugs
we've
had,
for
example,
in
the
and
the
in
the
unit
test,
we've
got
very
slow
unit
tests.
I
mean
the
unit
tests
on
the
sixty
eight
seconds,
we're
one
of
the
flakiest
test.
J
Also
and
the
other
part
that's
a
bit
bothering
you.
If
we
look
at
the
different
tests
that
set
up
and
the
tear
down
is
very
painful
and
I
think
so
and
I
think
it's
basically
a
sign
that
the
architecture,
ease
and
human
are
thick
and
it's
I,
don't
wanna
I
mean
this
is
the
point
of
this
document
is
to
give
an
idea
of
how
we
could
actually
break
the
architecture
down
in
multiple
modules,
and
so
this
is
the
current
architecture.
J
If
we
look
at
and
we've
got
basically
three
different
layers,
this
clothing
and
the
CM,
the
device
plug-in
handler,
which
is
in
the
qubit
/e
and
module
the
manager
which
is
in
the
device,
plug-in
module
and
the
income
and
so
and
most
of
the
architecture
and
the
architecture.
Documents
is
centered
around.
Basically
splitting
these
huge
monolithic
structures
and
basically
reduced
codecop
leg
and
the
call
the
complex
mutex
logic
that
we
have
in
it.
J
And
if
we,
if
we
look
at,
for
example,
the
last
wire
that
we've
seen,
we
basically
have
a
I
mean
we're
basically
having
mutex
is
calling
different
leaf
functions
and
we've
got
basically
huge,
critical
sections.
That
I
think
will
hurt
performance
in
the
long
term.
So
I
mean
you're,
chica
tech.
The
architecture
I
was
thinking
about
is
splaying,
basically
splitting
every
storage
responsibility
out
of
there
those
classes.
J
J
B
No,
no
first
of
all,
like
I,
think
I
must
appreciate
you
doing
this
work.
I
haven't
seen
any
other
feature
in
the
in
the
node
area,
having
gone
through
so
much
of
like
thorough
architectural
dogs.
So
thanks
a
lot
for
like
spearheading
this,
this
will
definitely
benefit
anyone
else,
who's
trying
to
contribute
in
this
area.
B
B
J
And
I
mean
keep
in
mind
that
the
whole
point
of
the
design
is
so
that
we
can
actually
have
two
PRS.
They
are
currently
good
examples
of
how
we
can
actually
get
from
a
module.
That's
currently
I
could
say
misbehaving
and
because,
like
is
I,
was
looking
at
the
facts,
for
example,
and
we're
probably
in
the
top
three
of
the
ficus,
a
unit
test
which
seems
a
bit
odd
for
unit
tests,
but
yeah.
B
B
I
would
say
is
like
in
in
your
mind
as
you're
developing
these
abstractions
just
have
a
conceptual
difference
between
interfaces
and
like
just
objects,
just
like
pure
Strax
and
go
go
a
line,
so
you
don't
necessarily
need
to
have
like
interfaces
for
every
single
level.
Each
interface
is
considered
an
abstraction
it
should
have.
It
should
make
some
sense
answers.
So
just
have
that
in
mind,
and
and
and
whenever
you
think
this
dog
is
ready
when
the
dog.
J
B
A
F
K
Can
hear
you,
okay,
I'm
a
sick,
eh
and
I'm
walking
on
stab
device
plug
in
for
a
tweeters
now
and
I
helped
post
here
is
a
proposal
on
github
and
received
some
comments
that
it
is
complicated,
so
I
update,
my
planning
on
github
and
I
will
implement
fake
device,
probably
in
first
so
this
is
my
first
time
to
contribute
to
community,
so
I
just
want
to
make
sure
everything.
I
do
is
right
for
community
and
comments
about
this
proposal.
Our
welcome
50
was
flooding.
K
Will
raid
dummy
device
file
from
temp
their
directory,
so
we
can
save
our
scene
for
into
this
file
and
we
can
inject
arrow
into
this
file.
It
is
deployed
by
team
instead,
so
we
can
do
easily
I
think
we
can
deploy
this
fake
device
plug
in
it.
We
know
the
test,
so
we
can't
deploy
our
port
that
will
request,
see
it's
fake
device,
but
a
fake
device
to
test
kubernetes
behaviors.
K
B
H
B
As
a
goal,
if
we
can
get
the
device
plug-in
tests
running
for
every
PR
and
like
running
as
part
of
my
regular
CI,
without
having
like
any
actual
hardware
differences,
that
is
going
to
be
super
helpful
and
as
a
next
step
once
we
as
a
community
think
that
the
device
plug-in
interface
is
reasonably
stable,
then
we
can
probably
like
request,
adding
it
to
the
conform
and
sweet
as
well.
Okay,.
A
D
A
A
A
B
It's
basically,
we
can
figure
out
a
way
to
make
the
deployment
actually
work
in
the
test
scenario
and
having
it
having
a
demo
plug-in
demo.
Plug-In
image
should
also
be
possible.
For
example,
we
we
can
place.
We
can
place
this
actual
binary
on
the
host
as
part
of
the
no
d3x
and
then
have
the
diamond,
such
as
execute
a
file
from
the
course.
A
It's
like
the
idea
that
being
separate
because
like
if
it's
so
closely
tied
to
our
ETS,
then
like
you
could
make
a
change
in
one
and
the
other.
At
the
same
time,
and
of
course
it's
like
if
we
somewhat
manager
a
dummy,
plugins
separate
from
our
normal
product
and
like
maybe,
we
won't
break
it
as
likely
or
it'll,
be
hard
to
it'll
be
like
to
PRS.
You
have
to
open,
but
take
separate
reviews
to
actually
get
a
change
in.
That
might
stop
you
from
breaking
something.
B
In
addition
to
that,
I
also
feel
like
having
having
a
dummy
plug-in
would
also
like
levers.
Will
integration
tests
rather
than
like
end-to-end
tests,
because
end-to-end
tests
are
expensive.
It's
like
for
many
reasons,
integration,
tester,
Chi,
so
I'm,
hoping
that
we
can
reuse
some
of
the
logic
for
integration
test
as
well.
H
E
B
A
Okay,
but
I
I
guess
we
can
talk
through
that,
but
only
speaking
like
it
I
would
prefer
if
we
built
like
an
artifact,
but
it
was
a
no
ops
device
plug-in
and
somehow
we're
able
to
use
that
in
our
testing,
but
like
not
do
them
double
T's.
But
if
I'm
the
only
one.
That's
strongly
thinking
that
way,
then
I
don't
want
to
hold
up
the
effort,
but
I
think.
B
So
the
wand
for
like
having
this
like
standard
implementations
that
you
used
to
verify
the
implement
doesn't
exist
as
of
now,
but
once
we
graduate
the
interface
out
of
alpha,
then
I
guess
that
will
become
a
necessity.
So
for
now
like
we
can,
you
can
just
make
sure
that
the
versions
we
came
to
support
are
actually
supported.
I
Device
plugins,
probably
a
good
fit
for
GPUs
and
such,
but
we
look
at
the
FPGA-
is
a
lot
more
stuff
which
need
to
be
added
for
that
if
pieces
of
programmable
devices,
so
the
current
AP
responder
and
look
at
programming
or
image
management,
matching
images
to
specific
devices
and
so
on.
The
other
aspect
is
that
lot
of
the
representation
of
images
and
accelerators
in
FPGA
world
is
not
standardized,
so
we
had
leave
lorath
of
device
plugin,
so
we
don't
necessarily
expose
all
other
than
the
api.
I
B
K
A
A
Alright,
with
that
I
will
stop,
recording
thanks
everyone,
and,
as
always,
if
you
have
topics
for
future
meetings,
just
feel
free
to
put
them
on
the
agenda
and
I
think
moving
forward.
If
I
don't
see,
topics
on
the
agenda
but
the
day
prior
I
will
cancel
the
following
days
meeting
so
maybe
I
motivate
thank.